Computer-implemented method for evaluating state of surge arrester

Abstract

To propose a method which can use collected data to automatically evaluate an operation condition so that further behaviors of a surge arrester can be predicted more accurately.SOLUTION: The present invention relates to a computer-implemented method for monitoring the current situation of a surge arrester and evaluating behaviors, the computer-implemented method including: providing a measured value of at least one surge arrester (S10); standardizing the provided measured value (S12); extracting a parameter for characterizing each surge arrester from the standardized measured value (S14); determining the state of each surge arrester by using a machine learning algorithm (S16); and outputting an operation recommendation for each surge arrester on the basis of the determined state of the surge arresters (S18).SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a method for assessing the condition of surge arresters based on their creepage and leakage currents. [Background technology]

[0002] Surge arresters, also called surge protection devices, are used in electrical energy transmission and distribution systems to protect devices connected to these systems from damage caused by surges due to lightning or switching operations in the network.

[0003] Typically, surge arresters use varistors to perform their function. Varistors can be zinc oxide components with non-linear electrical resistance characteristics. This means that their resistance decreases rapidly when a threshold is exceeded. This property allows surge arresters to rapidly and effectively divert surges to earth. Surge arresters are usually connected between the conductor and earth for this purpose. In case of overvoltage, the surge arrester safely diverts the excess energy to earth, thereby limiting the line voltage to a value that does not exceed the insulation value of the connected power technology, while isolating the line from earth at normal operating voltages.

[0004] Surge arresters are designed to operate below their maximum continuous operating voltage (Uc) during normal operation, which presents a high resistance and allows only small leakage currents to flow through the active parts of the surge arrester. Surge arresters are usually dimensioned so that the leakage current does not cause excessive power losses or exceed certain limits during normal operation.

[0005] However, short or continuous overloads on the surge arrester, poor design of the field control element or the use of zinc oxide varistors without long-term stability can lead to a degradation of the varistor performance, whereby the UI (voltage-current) characteristic curve of the surge arrester drops suddenly or even continuously and a higher leakage current flows thus to the new operating point. The power loss is radiated from the surge arrester housing in the form of thermal energy. If the electrical energy introduced into the arrester by the continuously flowing leakage current and by additional transient overvoltages exceeds the thermal energy that can be radiated outwards from the arrester, the arrester becomes thermally unstable and fails electrically. This is a rare failure.

[0006] Most surge arrester failures are due to moisture ingress. This usually occurs either through a faulty sealing system in surge arresters with gas capacity, or directly through the interface between different materials, as in the case of direct-cast polymer arresters or other arresters without gas capacity. Possible causes of such failures include, for example, errors in the manufacturing process, material fatigue or design weaknesses. In the case of surge arresters with sealed gas capacity, it is a lengthy process where moisture usually enters through leakage due to alternating pressure difference between the inside and outside of the surge arrester. When the temperature inside the enclosure falls below the dew point, moisture condenses along the active parts or inside the surge arrester housing. This can lead to flashover and the formation of creepage currents inside the housing, which can eventually cause a complete short circuit inside the housing and thus a sudden failure of the arrester.

[0007] Another common failure mode is the deterioration of the insulation of the arrester housing caused by surface contamination in the form of increased surface conductivity, tracking and erosion. This is a common problem with surge arresters used in so-called polluted areas. These surge arresters show the accumulation of conductive precipitates after a certain time. This causes creepage currents along the surface of the housing and initially partial flashovers, which can damage the surge arrester housing or impair the hydrophobicity of silicone insulated surge arresters. In addition, only partially conductive housing sections can cause high radial fields between the varistor column and the outer layer of contamination, causing partial discharges and thus damage inside the housing. If this contamination is detected early, it can be remedied by cleaning the surface. In some cases, the contamination can lead to external flashovers and permanent damage to the surge arrester.

[0008] In view of the many different causes of failures in surge arresters, power system operators are faced with the challenge of dealing with aging equipment. As a result, the topic of "asset life cycle management" has become a hot topic in the energy sector. To meet this challenge, manufacturers of power engineering equipment offer monitoring systems specially developed for their products. For example, patent US 5,999,433 discloses such a monitoring system for surge arresters. A similar concept is described in https: / / p3.aprimocdn.net / siemensenergy / fb8f300d-debc-415c-87e6-b03b01315915 / Sensarrester-flyer-with-Frost-Sullivan-Award-pdf_Original%20file.pdf.

[0009] These monitoring systems monitor a variety of devices such as transformers, switches, disconnectors, converters, circuits, isolators and surge arresters, resulting in a large amount of data. However, this data is usually very abstract, so it is necessary to have experts in each field of application to evaluate it.

[0010] The large amount of data and experience, especially measurement data from surge arresters, requires careful evaluation as the information content is not immediately apparent. As a result, many operators of power systems are unable to assess the condition or remaining life of such devices.

[0011] To make matters worse, power system operators usually use a large number of surge arresters of different types, from different years of manufacture and from different manufacturers. An overview of the various basic designs can be found, for example, at https: / / cache.industry.siemens.com / dl / files / 132 / 109747132 / att_918329 / v1 / Metalloxid-Ableiter_in_Hochspannungsnetzen_-_Grundlagen.pdf.

[0012] Well-known management schemes require that alarms be built into existing systems to send notifications when the typical leakage current limits of surge arresters are exceeded. However, for complex signal patterns that could allow predictions about future failures, operators must contact an expert to request a manual assessment of the surge arrester's condition. [Prior art documents] [Patent documents]

[0013] [Patent Document 1] German Patent No. 102015013433 (DE10 2015 013 433.7) [Non-patent literature]

[0014] [Non-Patent Document 1] https: / / p3.aprimocdn.net / siemensenergy / fb8f300d-debc-415c-87e6-b03b01315915 / Sensarrester-flyer-with-Frost-Sullivan-Award-pdf_Original%20file.pdf [Non-Patent Document 2] https: / / cache.industry.siemens.com / dl / files / 132 / 109747132 / att_918329 / v1 / Metalloxid-Ableiter_in_Hochspannungsnetzen_-_Grundlagen.pdf Summary of the Invention [Problem to be solved by the invention]

[0015] In light of this, the object of the present invention is to propose a method, whereby the collected data can be used to automatically assess the operating conditions in order to better predict the further behavior of surge arresters, to service or replace them in good time, and to prevent damage or sudden failures due to defects in the surge arresters. [Means for solving the problem]

[0016] The task is solved by the subject matter of the independent claims.

[0017] According to a first aspect of the invention, this task is solved by a computer-implemented method for determining the operating state of a surge arrester, the method comprising: providing a measurement of at least one surge arrester; normalizing the provided measurements; extracting parameters for characterizing the at least one surge arrester from the standardized measurements; determining a status of at least one surge arrester using a machine learning algorithm; outputting a recommendation for operating at least one surge arrester; Equipped with.

[0018] The measured values, such as the peak value of the leakage current and / or the resistive current based on the third harmonic of the leakage current, can preferably be segmented time series determined in the surge arrester using a corresponding sensor, in particular a current measuring device. The data thus obtained in the measured time series can be particularly preferably segmented in suitable windows and used later for window-based analysis.

[0019] Measurement values ​​of different surge arresters can have different orders of magnitude, this depends on the surge arrester being monitored and the respective use case. The signals are standardized for better explainability and comparability of the data. For example, the z-scope standardization method can be used for standardization, which converts the measurements into a distribution with a mean value of 0 and a standard deviation of 1. The dynamic behavior of the data is preserved and the units become dimensionless.

[0020] In some cases, normalization may precede standardization. The main difference between data normalization and data standardization is that data normalization brings data to a particular scale. The goal of this method is to adjust the data so that they have the same range of values. This makes it possible to compare data with different scales or units.

[0021] Standardization, on the other hand, transforms the data into a standard normal distribution. It removes the mean and scales the data to unit variance. The goal of standardization is to bring the data into a standardized form that allows the range of the data to be preserved and the distribution of values ​​to be the same. This makes measurements of different surge arresters comparable.

[0022] Then, characteristic quantities are extracted from the standardized data and used to evaluate the behavior of the surge arrester. With the help of various feature extraction methods, a large number of features can be easily obtained, which have advantages and disadvantages. On the one hand, the large number of used features increases the precision and depth of the analysis. On the other hand, various features often contain redundant information that can have a negative effect on the learning of the learning algorithm. This can make the model unstable.

[0023] Depending on the different patterns in the measurements, various features can be extracted from the measurements. The patterns found in the measurements can be roughly divided into weather-related signal patterns, grid-related patterns and surge arrester malfunction patterns. The weather-related and grid-related patterns are not relevant for assessing the arrester status.

[0024] In case of a network interruption due to a fault or maintenance work, the leakage current value usually shows a value close to 0 μA. If an earth fault occurs in a network with isolated zero conductors (star connection), the affected line is pulled almost to earth potential, causing a significant drop in the voltage at the surge arrester and a reduction in the leakage current.

[0025] The other two phases of the network take on chained voltage values ​​that lead to an increase in leakage current, which can for example also be observed in meteorological signal patterns. Surge arresters are usually installed outdoors and are subject to different weather conditions such as rain, fog, humidity and solar radiation. Both phenomena show similar patterns in the measurements, but they are independent of the arrester state and must be separated from the user data before feature extraction.

[0026] A malfunctioning surge arrester will show different signal patterns depending on the type and severity of the error. Malfunctions can be subdivided, for example, into a strong periodically pulsed signal pattern, a very strong stochastically fluctuating signal pattern, or a trend increase in the signal.

[0027] The main objective of characteristic extraction is to replace a large amount of data with a few meaningful extracted characteristics that contain all the information necessary to solve the problem under investigation. Therefore, the characteristics that best characterize the surge arrester are extracted from the available database.

[0028] Once the parameters for the surge arrester are determined, the state of the surge arrester is determined from them. Machine learning algorithms are used for this purpose.

[0029] Machine learning algorithms can take many forms, such as linear models, decision trees, support vector machines, neural networks, and many others. Machine learning algorithms are optimized by learning from training data, recognizing signal patterns and rules to make the best possible predictions or classifications for new data.

[0030] The effectiveness of a machine learning algorithm depends on a variety of factors, including the quality and quantity of training data, the selection of the model, the model configuration, and the evaluation of the model using evaluation metrics. Models can be continuously improved and optimized to maximize accuracy and performance.

[0031] A linear regression model assumes a linear relationship between a dependent variable and one or more independent variables and is used to make predictions for continuous values.

[0032] A Support Vector Machine (SVM) is a model often used for classification or regression to detect signal patterns in data. SVMs try to find the optimal separation between different classes or fit a continuous function to the data.

[0033] A decision tree is a model that produces decision rules in the form of a tree diagram. Decision trees are used to split data into features to enable prediction or classification.

[0034] Naive Bayes is a probabilistic model based on Bayes' theorem used for classification. Naive Bayes assumes that features are independent of each other and calculates the probability of a particular class based on given features.

[0035] Neural networks are models preferably used to process images or other matrix-based data. The architecture of a neural network comprises several knots, neurons or nodes arranged in layers. Several layers may be designed as sub-networks. Sub-networks can also share one or several layers to solve their tasks with the same output values. Then each sub-network can have its own layer that works only to solve the specific task of the sub-network.

[0036] Preferably, a machine learning algorithm based on a clustering method is used to carry out the proposed procedure. In clustering, the data is divided into groups that are most similar to each other. In particular, this can be a model that performs a partitioning cluster analysis. The aim of this is to divide the data of the training data set into several groups that are initially determined. This division is iteratively optimized until a target function, such as the mean squared error, reaches an optimum. Common methods of this form are, for example, the K-means, K-medoids and CLARANS procedures.

[0037] Recommendations for further operation of the surge arrester can be derived from the surge arrester's condition. For example, a surge arrester categorized into a class representing severe weathering can be classified as not very reliable. The operator can then maintain, repair or replace the surge arrester in order to prevent damage or unscheduled failure of the surge arrester. In another case, the surge arrester can be categorized, for example, into a clear error class, in which the severity of the error is only minor and therefore the probability of failure is also low. This does not lead to any direct action for the operator and the surge arrester can continue to operate.

[0038] Preferably, surge arresters may be divided into different classes assigned to weather-related conditions, service-related conditions and fault conditions. In an embodiment, further subclasses may be used to describe the extent of each condition in more detail.

[0039] In one embodiment, normalization of the provided measurements is performed after smoothing of the provided measurements.

[0040] Smoothing the measurements has the beneficial effect of minimizing the influence of extreme values ​​on the remaining measurements.

[0041] In one embodiment, the normalization of the provided measurement values is performed after removal of the "off state" from the provided measurement values.

[0042] The off state refers to a period during which the line is powered off or a ground fault occurs, causing the leakage current to be approximately zero. These low values can be identified and removed from the measurement values, for example, using the so-called "box plot" method.

[0043] No information regarding the state of the surge arrester can be obtained from the off-state measurement values. Therefore, reducing the measurement values around the off state reduces the computational effort of the machine learning algorithm and increases the efficiency with which the characteristics are determined.

[0044] In one embodiment, the measurement data is acquired by measuring the leakage current in the surge arrester, whereby the peak current and the resistive current are determined from the leakage current.

[0045] The leakage current in the surge arrester consists of a sinusoidal capacitive component that is phase-shifted by -90° with respect to the voltage signal and a resistive component that is in phase with the voltage signal and in the form of a periodic pulse signal.

[0046] These two components combine to form the total leakage current, and two important core values can be extracted from the total leakage current. The peak value of the leakage current and the third harmonic with respect to the main frequency of 50 or 60 Hz. The peak value of the leakage current is always determined by the dominant capacitive or resistive current component. At low voltage levels (U < Uc), the peak current takes the peak value of the capacitive component. At higher voltage levels, mainly above the rated voltage (Ur), the peak current is based on the peak value of the resistive component.

[0047] The resistive current is particularly meaningful for evaluating surge arresters with respect to the capacitive current, since the resistive current changes significantly when the current-voltage characteristic of the surge arrester degrades; in such cases, the capacitive current does not change significantly.

[0048] However, for optimal monitoring of the surge arrester condition, it is advantageous to monitor both components of the leakage current in the surge arrester.

[0049] One method for measuring the leakage current is based on the fact that the nonlinear current-voltage characteristic of the surge arrester generates harmonics in the leakage current. The proportion of harmonics in the leakage current depends heavily on the peak value and the operating point of the resistive current and thus varies with the voltage and temperature of the surge arrester. Experience has shown that the resistive current can be accounted for by the third harmonic using a factor. For example, the total leakage current can be obtained periodically, approximately once per hour, from measurements over 10 wavelengths at normal line voltages of 50 or 60 Hz.

[0050] For resistive current evaluation, the third harmonic is the most common. It offers the best measurement sensitivity.

[0051] The leakage current with its peak current and resistive current components advantageously forms a measurable variable that is relatively easy to determine and from which good characteristic variables can be derived.

[0052] In one embodiment, extracting the features involves converting the standardized measurements into a frequency spectrum, where the features include discrete spectral components from the frequency spectrum and trends in the frequency spectrum.

[0053] Waveform data can be analyzed in the time domain, the frequency domain, and the time-frequency domain, each of which provides different insights and allows different aspects of the data to be examined.

[0054] Time domain analysis considers measurements directly in their time course. Signal behavior is examined based on parameters such as signal patterns, signal trends, periodicity, and other important signal properties such as stationarity or non-stationarity. Stationary signals are characterized by invariant statistical properties such as mean, variance, and autocorrelation over time, whereas the statistical properties of non-stationary signals change over time.

[0055] Many time series analysis methods are only suitable for stationary measurement series. However, the signals analyzed using the proposed method may not be stationary since the monitored surge arresters are subject to many external influences that affect their signal patterns.

[0056] Non-stationary signals can be treated as stationary by windowing. Special procedures can be applied to convert the signal into a stationary signal. Time domain analysis is particularly suitable for detecting signal patterns or changes in the dynamic behavior of a measurement sequence over time. For example, it is possible to identify anomalies or threshold changes.

[0057] For example, the behavior of a surge arrester is influenced by the course of the day, i.e. the temperature difference between day and night.

[0058] Frequency domain analysis examines the frequency content of a time series by breaking the signal down into its spectral components. This involves analysing the amplitude, phase relationships and dominant frequencies of the signal.

[0059] There are methods for transforming signals from the time domain to the frequency domain and vice versa. The fast Fourier transform (FFT) is a commonly used method. This method allows the frequency specific energy and phase dynamics of the signal to be examined or can be used as an adjunct to various filters or convolution algorithms.

[0060] FFT is well suited for stationary signals since the relevant characteristics of the signal can be derived in the time domain. Non-stationary signals have power spectra that can be difficult to interpret, for which the classical FFT function is severely limited. On the other hand, the Welch method divides the non-stationary signal into successive windows, performs a Fourier transform for each window, and averages the power spectrum.

[0061] Due to the fact that measurements of different surge arresters may contain complex signals whose dynamic behavior changes completely over time due to occurring fault conditions or weather effects, an exclusive consideration in the time domain is not always sufficient. Therefore, to capture this entire dynamic behavior and, as a consequence, its changes, methods can be implemented. FFT and DWT are preferred methods to map this dynamic behavior in the frequency domain.

[0062] In particular, 1 / day and / or 2 / day or 1 week discrete spectral components of peak current can be used to characterize the surge arrester. Seasonal observations over longer periods are also possible.

[0063] The emphasis on the 1 / T spectral component points to a connection with daily processes, e.g., morning condensation and the resulting high surface currents. Surface currents are regularly caused by moisture and increase with a dirt layer. In addition, normal operating conditions may also be characterized by a 1 / T behavior, since line voltages usually follow a similar load-dependent daily course.

[0064] The 2 / day discrete spectral components in the peak current may indicate a correlation with daily processes, such as morning soiling and condensation, and evening condensation and resulting high surface currents.

[0065] Moreover, higher harmonics of the discrete spectral components of the peak currents can be used. These higher harmonics can represent seasonally repeating behavior, for example over the course of a day. In the absence of noise in the higher spectral range, and thus of stochastic behavior, in other words 1 / f behavior, this can be interpreted as the presence of a repeating environmental effect linked to the time of day, such as heavy rainfall in monsoon regions or similar.

[0066] In one embodiment, a comparison is made between the signal energy in one or more low frequency ranges of a measurement series and one or more higher frequency ranges, which is used to determine whether stochastic behavior is present.

[0067] In one embodiment, the characteristics also include the signal to noise ratio in the frequency spectrum, and in particular in a defined section of the frequency spectrum.

[0068] In one embodiment, the parameters also include a trend in the frequency spectrum.

[0069] If the signal is subject to stochastic noise and all spectral components are weighted equally, a tilt in the spectrum in the range m ≈ 0 can occur. This behavior can be observed especially in surge arresters with moisture inside, since the high current flow creates small conductive paths that dry out quickly and the evaporated water condenses elsewhere.

[0070] In one embodiment, the parameters also include trends in the time domain.

[0071] An upward trend in the time domain m>0 indicates an intensifying dynamic in the signal, or continuously increasing peak and resistive currents. All types of faults can appear in the time domain with an upward trend. An upward trend is therefore an indication of the presence of a fault. However, this characteristic is not very suitable for identifying a specific fault. However, an exception to this could be the advanced aging of a surge arrester, since in this case the values ​​increase continuously and the signal has little energy.

[0072] In one embodiment, the characteristic further includes a correlation value between the peak current and the resistive current in the time domain.

[0073] The inventors have experimentally determined that the correlation between peak current and resistive current, as well as the analysis of these two signals in the frequency domain, is useful in identifying the condition of a surge arrester.

[0074] The correlation between the peak current and the resistive current is high when strong surface currents are present. The resistive component of the leakage current, extracted from the third harmonic leakage current component, increases equally because the harmonic components are added to the current by the short dry band discharge. However, this also applies to internal currents caused by moisture ingress that flash over partially dry surfaces. A partial short circuit resulting from the varistor block generates a high peak in the resistive current that exceeds the capacitive current component and thus determines the peak current. A strong correlation between the resistive current and the peak current, especially greater than 0.5, may indicate heavy contamination or heavy moisture ingress. If the correlation is low, especially close to 0, it may be concluded that there is no or only slight surface contamination or no malfunction.

[0075] A determination of the correlation coefficient between the peak current and the resistive current in a segmented window may be used, for example the Pearson correlation coefficient may be used for this.

[0076] The described properties may preferably be used together and / or in groups.

[0077] In one embodiment, the leakage current is determined during operation of the surge arrester.

[0078] This advantageously enables the surge arrester to be checked for malfunctions, preferably at regular intervals, and treated, repaired or replaced depending on the assessment. For this purpose, the surge arrester may be equipped with a device for transmitting the measured values ​​to a central evaluation system.

[0079] In a further embodiment, the measurement values ​​can be collected over a defined time period and then evaluated in bundle form, which means that the continuous transmission of the measurement values ​​to the central evaluation system can be limited to a few transmissions, for example supported by the transport of data carriers.

[0080] In a further aspect, the invention relates to a computer program having a program code for performing the method described above, when the computer program runs on a computer.

[0081] In a further aspect the invention relates to a computer readable data carrier carrying a program code of a computer program for performing the method described above, when the computer program is run on a computer.

[0082] In a further aspect, the invention relates to a system for evaluating the behavior of a surge arrester, which system is designed to carry out the method described above.

[0083] For example, the system may be implemented as a web server with a web application, as a local computing system, or as a portable device for checking surge arresters in the field.

[0084] The described embodiments and further developments can be combined with one another in any way.

[0085] The accompanying drawings are intended to provide a further understanding of the embodiments of the present invention, and together with the description, serve to explain the principles and concepts of the invention. [Brief description of the drawings]

[0086] [Figure 1] 2 shows a schematic diagram of a method sequence according to an embodiment; [Diagram 2] 5 shows a schematic diagram of a method sequence according to a further embodiment; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0087] FIG. 1 shows a simplified sequence of a method according to one embodiment.

[0088] The method starts in step S10 by providing measurements of the surge arrester. The measurements can be continuously recorded and evaluated or collected and evaluated in batches. For example, the measurements can be transmitted from the surge arrester via a network, in particular a wireless network, to a data processing system. The data processing system can be at the operator's or at the manufacturer's of the power engineering system. The primary measured values ​​are current measurements with a resolution well below the grid frequency. These primary measured values ​​are collected and stored over a sufficiently long period of time, for example over 10 wave sequences. The primary measured values ​​can be obtained continuously, but it is preferable to measure these values ​​at periodic intervals such that a set of primary measured values ​​is obtained approximately once every hour or once every 15 minutes.

[0089] The set of primary measurements are then used, preferably in the measurement device itself, to obtain the peak current, the third harmonic of the leakage current (hereinafter often referred to as "leakage current" for simplicity), and, if applicable, further characteristic or derived measurements. When the term "measurement" is used below, it refers uniformly to one of the derived measurements, unless the primary measurement is explicitly referred to.

[0090] In a second step S12, the provided measurements are standardized. The aim of the standardization is to make the measurements of the surge arrester under investigation equivalent to the whole set of surge arresters used to train the machine learning algorithm. In this process, the measurements are corrected for effects resulting from the special configuration of the surge arrester being analyzed. The correction can take into account, for example, the type of surge arrester, its use or the infrastructure surrounding it. In particular, the line voltage with respect to the arrester rated voltage, the ratio of the rated voltage to the reference voltage of the surge arrester, the type of varistor used, the manufacturing tolerances in the ZnO deposition process during the fabrication of the surge arrester, the grounding conditions and / or the capacitive boundary conditions in use can be taken into account. Furthermore, the measurements can be standardized with respect to the surge arrester housing made of ceramic or plastic, the shielding design and / or position and its climatic conditions.

[0091] In step S14, characteristic quantities are derived from the standardized measured values. The characteristic quantities serve to recognize the state of the surge arrester. Various features arising from the behavior of the surge arrester can be used as characteristic quantities. In particular, the leakage current (peak value or third harmonic) can be used as a measured value and the features determined therefrom can be used as characteristic quantities.

[0092] The parameters may include, but are not limited to, 1 / day discrete spectral component in the peak current, 2 / day discrete spectral component, higher harmonics of the discrete spectral component, a decreasing trend (mf<0) in the discrete frequency spectrum of the peak current, a signal-to-noise ratio in the frequency spectrum, an increasing trend (mf>0) in the discrete frequency spectrum, a correlation between the peak current and resistive current in the time domain and / or a trend mz in the time domain of the peak current and the peak current.

[0093] In step S16, a machine learning algorithm is used to determine the state of the surge arrester. To do this, the characteristic variables are input as input variables to the machine learning algorithm. The machine learning algorithm is trained to recognize the state of the surge arrester using training data. The training data includes characteristic variables of a plurality of surge arresters in different situations and with different states.

[0094] In particular, the classification task may relate to a state determination in which a machine learning algorithm assigns a surge arrester to one of several classes based on its characteristic variables, each class representing a state.

[0095] How the assignment is done depends on the selected model and its architecture. Different models are constructed differently. In principle, any model that is suitable for classification tasks can be used. In this case, clustering is performed since no labeled data is available. The result is a group formation, whereby further procedures (for classification) can be applied later.

[0096] In a final step S18, the state and a representation of the state are reported together with a corresponding recommendation for action or together with an alert.

[0097] In one embodiment, a surge arrester can be assigned to several states. For example, a first state can indicate the degree of contamination and a second state can indicate the condition of the seal in the surge arrester. The properties "housing dirty / not dirty" and "seal intact / broken" are not mutually exclusive. Moreover, the states can be output as parameters in a spectrum, whereby, for example, the properties of the surge arrester are quantified.

[0098] According to the invention, additional information may be obtained from a synopsis of the measurements or characteristic parameters of a group of surge arresters installed in the immediate vicinity and each of them assigned to one of the three phases of the network.

[0099] Figure 2 shows a variation of the method from Figure 1. The process flow of the method according to a further embodiment, as shown diagrammatically in Figure 2, also starts with the provision of derived measurements of the surge arrester in step S10.

[0100] Prior to normalization, the measurements are pre-processed in step S20. In particular, pre-processing can include segmenting the measurements to separate individual signals. For example, peak currents and resistive currents can be determined from the measured leakage currents. Furthermore, the measurements can be filtered, for example to remove noise of the measurement device or other extraneous noise.

[0101] After standardization in step S12, characteristic quantities are extracted in steps S14a and S14b. To do this, the standardized measurements are transformed into frequency space. Characteristic quantities can then be extracted from the measurements in the time domain in step S14a and from the frequency spectrum in step S14b.

[0102] In this embodiment, the determination of the status in step S16 and the output of recommendations for action in step S18 may proceed as described with respect to Figure 1. However, following the evaluation in step S18, a validation of the evaluation may be performed in step S22.

[0103] The assessment can be verified, for example, by checking whether the conditions assigned to the surge arrester are actually present when the surge arrester is processed as a result of the assessment. For example, the degree of contamination of the housing or the humidity inside the surge arrester can be determined. Additionally or alternatively, samples can be taken to check the processing of the measurements with machine learning algorithms.

[0104] Alternatively, one of the surge arresters that has been replaced by the method of the present invention may be returned to the manufacturer for detailed laboratory testing.

[0105] The results of these studies can then be fed back to train the system.

[0106] While current monitoring devices only use the primary measurement values ​​of the surge arresters to output a simple "red-yellow-green" signal, by using machine learning and taking into account a large amount of measurement data from a wide variety of surge arresters in a wide variety of environments, the invention allows much more accurate and reliable statements to be made about the condition of each individual surge arrester, without the need for experts to carry out complex individual assessments themselves. [Explanation of symbols]

[0107] S10 Provide measurements S12 Standardize measurements S14 Extracting characteristic parameters S14a Extracting characteristics from time periods S14b Extracting characteristics from frequency spectra S16 Determining status S18 Derive recommendations for action S20 Preprocessing measurements S22 Verifying the evaluation

Claims

1. 1. A computer-implemented method for evaluating a performance of a surge arrester, comprising: Providing measurements of one or more surge arresters (S10); A step (S12) of standardizing the provided measurements; Extracting parameters for characterizing at least one surge arrester from the standardized measurements (S14); Determining a status of the at least one surge arrester using a machine learning algorithm (S16); outputting a recommendation for action regarding operation of said at least one surge arrester (S18); A computer-implemented method comprising:

2. The computer-implemented method of claim 1 , wherein the standardization (S12) of the provided measurements is performed after smoothing of the provided measurements.

3. The computer-implemented method of claim 1 or 2, wherein the normalization (S12) of the provided measurements is performed after removal of off states from the provided measurements.

4. 4. The computer-implemented method of claim 1, wherein the measurements are detected by measuring a leakage current in the surge arrester, and a peak current and a resistive leakage current are determined from the leakage current.

5. The computer-implemented method of claim 1 , wherein the measurements are determined during operation of the surge arrester.

6. 6. The computer-implemented method of claim 1, wherein the extraction (S14) of the characteristic quantity comprises a transformation of the standardized measurement value into a frequency spectrum, and the characteristic quantity comprises discrete spectral components from the frequency spectrum and / or a trend in the frequency spectrum of the peak current.

7. The computer-implemented method of claim 6 , wherein the characteristic quantity further comprises a signal-to-noise ratio in the frequency spectrum, in particular in a defined section of the frequency spectrum.

8. The computer-implemented method of claim 1 , wherein the characteristic quantity further comprises a correlation value between the peak current and a resistive current in the time domain.

9. A computer program comprising a program code for performing the method according to any one of claims 1 to 8, when the computer program is executed on a computer.

10. A system for issuing recommendations for the operation of a surge arrester, the system being configured to carry out the method of any one of claims 1 to 8.

Images