Transformer oil chromatographic trend monitoring method based on multi-scale analysis

Abstract

The application provides a transformer oil chromatographic trend monitoring method based on multi-scale analysis, comprising the following steps: S1, collecting time series data G(t) of gas concentration; S2, performing multi-scale sliding window segmentation fitting on the data G(t) to obtain segmented fitting data; S3, performing clustering processing on the segmented fitting data and converting the segmented fitting data into a symbolic sequence Σ(t); S4, constructing a three-dimensional feature vector based on the symbolic sequence and a concentration change rate R(t) and a second derivative A(t) obtained from G(t), and calculating a support degree sup and a confidence degree conf; S5, identifying abnormal candidate segments according to the calculation results and correcting the abnormal candidate segments to obtain corrected data; and S6, performing trend modeling and prediction on the corrected data, outputting a trend vector T(t+Δt), and performing hierarchical diagnosis. The application is helpful to reduce noise interference and improve the accuracy of abnormal identification.

Description

Technical Field

[0001] This application relates to the field of power equipment condition monitoring and diagnosis technology, and in particular to a method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis. Background Technology

[0002] Transformers are critical equipment in power systems, and their long-term stable operation directly affects the safety and reliability of the power grid. Due to electrical, thermal, and mechanical stresses, the internal insulation materials of transformers are prone to deterioration and the generation of decomposition gases. Dissolved Gas Analysis (DGA), a commonly used diagnostic tool, can detect the types and concentrations of characteristic gases dissolved in insulating oil to reflect the presence of latent defects in the equipment. In recent years, with the development of smart grids, online DGA monitoring devices have been gradually applied in operating fields, enabling continuous monitoring of dissolved gases and providing data support for condition assessment and maintenance decisions.

[0003] Existing online oil chromatography monitoring methods mostly rely on a single indicator for judgment, such as the ratio of hydrogen concentration, carbon monoxide concentration, or hydrocarbon gases, combined with a fixed threshold to infer the type of fault. When the concentration of a certain gas exceeds the set threshold, it is judged as an abnormal state. Although this type of method is simple, it mainly focuses on the instantaneous level of concentration and fails to reflect the evolution of the gas over time. At the same time, online monitoring data is easily affected by sampling fluctuations, sensor accuracy, and environmental interference. Without further trend modeling and data processing, the judgment results are often distorted or there may be missed or false alarms.

[0004] Therefore, existing technologies still have shortcomings in identifying abnormal patterns: they rely too heavily on static thresholds, fail to fully utilize the dynamic changes in gas concentration, and struggle to capture potential deterioration trends in a timely manner. These limitations weaken the reliability and predictability of monitoring results, leading to delays or biases in anomaly detection. How to more effectively utilize time-series information during online oil chromatography monitoring and establish trend analysis methods based on this information to improve the accuracy of anomaly identification and diagnosis has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a transformer oil chromatographic trend monitoring method based on multi-scale analysis to solve the problem that existing technologies rely on static thresholds and are difficult to effectively utilize the characteristics of gas concentration changes over time, resulting in inaccurate anomaly identification.

[0006] This application provides a method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis, including the following steps: S1, Collect time series data G(t) of gas concentration output by the online transformer oil chromatography monitoring device; wherein the characteristic gases corresponding to the time series data include hydrogen, methane, acetylene, ethylene, carbon monoxide and carbon dioxide; S2, perform multi-scale sliding window piecewise fitting on the time series data G(t) of the gas concentration to obtain piecewise fitted data; S3, perform clustering processing on the segmented fitting data, and convert the clustering results into a symbolic sequence Σ(t); S4. Based on the symbolic sequence Σ(t) and the concentration change rate R(t) and the second derivative of concentration A(t) calculated from G(t), construct a three-dimensional feature vector <G, R, A> composed of gas concentration, concentration change rate and second derivative of concentration, and calculate the support sup and confidence conf. S5. Based on the calculation results of the support (sup) and confidence (conf), identify anomalous candidate segments, and perform data correction on the anomalous candidate segments to obtain corrected gas concentration data. ; S6, the corrected gas concentration data Perform trend modeling and prediction, output trend vector T(t+Δt), and perform hierarchical diagnosis based on the trend results.

[0007] In one optional embodiment, the time series data G(t) of the gas concentration is obtained by continuously collecting the concentrations of hydrogen, methane, acetylene, ethylene, carbon monoxide and carbon dioxide output by the online oil chromatography monitoring device at preset time intervals Δt, wherein Δt is 1 to 10 minutes.

[0008] In one optional implementation, the multi-scale sliding window piecewise fitting is achieved by the following steps: S21, divide the time series data G(t) of gas concentration according to the short-time scale window length L1 and the long-time scale window length L2, and set the overlap rate r between adjacent windows; S22, within each window, a fitting operation is performed on the divided gas concentration data. The fitting operation includes establishing a linear function or a polynomial function and extracting the corresponding parameter vector from the function. S23, use the set of parameter vectors corresponding to each window as piecewise fitting data.

[0009] In one optional implementation, the short-timescale window length L1 ranges from 50 to 100 sampling points, the long-timescale window length L2 ranges from 300 to 500 sampling points, and the overlap rate r between adjacent windows is 20% to 40%.

[0010] In one optional implementation, step S3 includes the following steps: S31, use the segmented fitting data as clustering input; S52, the k-means clustering algorithm is used to classify the segmented fitted data; S53, determine the number of clusters k based on the maximization result of the silhouette coefficient Sc(k); S54. Assign symbolic identifiers to the cluster output categories and arrange them in chronological order to generate a symbolic sequence Σ(t).

[0011] In one optional implementation, the construction of the three-dimensional feature vectors <G, R, A> in step S4 includes: The time series data of gas concentration G(t) at sampling time t i The gas concentration value G(t) i ) as the G component in the three-dimensional feature vector <G, R, A>; The rate of concentration change R(t) is calculated according to the following formula. i ): ; Wherein, R(t) i ) indicates at time t i The rate of concentration change; The second derivative of the concentration, A(t), is calculated using the following formula. i ): ; Where A(t) i ) indicates at time t i The rate of change of the concentration; G(t) i ), R(t) i ) and A(t i The three-dimensional feature vectors <G, R, A> are combined and mapped to the corresponding segments of the symbolic sequence Σ(t).

[0012] In an optional implementation, the calculation method for support (sup) and confidence (conf) in step S4 includes: The three-dimensional feature vectors <G, R, A> are discretized according to the segmentation result of the symbolic sequence Σ(t) to obtain the itemset sequence; In the itemset sequence, count the number of times any itemset X and itemset Y appear simultaneously, count(X∪Y), and count the number of times itemset X appears, count(X). Calculate the support and confidence scores according to the association rule formula: ; ; Where N represents the total number of items in the itemset sequence, X and Y represent itemsets in the symbolic sequence Σ(t), count(X∪Y) represents the number of times itemset X and itemset Y appear simultaneously, and count(X) represents the number of times itemset X appears.

[0013] In one optional implementation, step S5 includes the following method for determining abnormal candidate segments: The calculated support score (sup) and confidence score (conf) are compared with the baseline support score (sup0) and baseline confidence score (conf0) obtained based on historical normal operation data. A segment is considered an abnormal candidate segment when one of the following conditions is met: or ; Where, sup0 represents the baseline support value obtained based on historical normal operation data, conf0 represents the baseline confidence value obtained based on historical normal operation data, α represents the bias judgment threshold of support, which ranges from 0.05 to 0.20, and β represents the bias judgment threshold of confidence, which ranges from 0.05 to 0.20.

[0014] In one optional implementation, the data correction method in step S5 is specifically as follows: The gas concentration data identified as abnormal candidate segments are filtered using Kalman filtering to reduce the impact of random noise on the data. After filtering, regression compensation is performed on the residual values ​​of the abnormal candidate segments to ensure the corrected gas concentration data. The data trend remains continuous with that of adjacent normal segments; When the corrected gas concentration data Reference data G of adjacent normal segments ref (t i The mean square error (MSE) of a given condition is satisfied when: ; The correction is deemed valid; wherein, G ref (t i ) represents the reference sequence, n represents the number of sample points involved in the calculation, and γ represents the preset error threshold, which is set to 10. -3 ~10 -2 .

[0015] In one optional implementation, step S6 specifically involves: For the corrected gas concentration data A time series forecasting model is used for modeling and forecasting to obtain the trend vector T(t+Δt); the time series forecasting model adopts the autoregressive moving average model ARIMA. The gas concentration growth rate r is calculated based on the trend vector T(t+Δt). g ; gas concentration growth rate r g The unit is % / day; When r g When the risk level is less than the threshold α1, it is classified as a Level 1 risk with corresponding low risk. When r g When the risk level is between α1 and β1, it is classified as a secondary risk within the medium-risk category. When r g When β1 is greater than or equal to β1, it is classified as a level 3 risk corresponding to high risk; Among them, the value of α1 ranges from 0.5% to 1.0% per day, and the value of β1 ranges from 1.0% to 2.0% per day.

[0016] Compared with the prior art, this application has the following beneficial effects: 1. This application provides a method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis. This application uses time-series gas concentration data G(t) as a basis, moving beyond single-point concentration values ​​or static ratios. Instead, it obtains piecewise fitted data through multi-scale sliding window piecewise fitting. The introduction of multi-scale analysis helps to simultaneously reflect long-term trends and short-term fluctuations. Piecewise fitting divides the continuous sequence into several segments, expressing the change characteristics of that time period through the overall trend within each segment. This not only reflects the dynamic characteristics of gas concentration over time but also facilitates comparison between different segments, providing a consistent input for subsequent processing.

[0017] 2. After obtaining the piecewise fitted data, this application further employs clustering to generate a symbolic sequence Σ(t). Based on this, a three-dimensional feature vector <G, R, A> composed of gas concentration, concentration change rate, and second derivative of concentration is constructed. Subsequently, the support (sup) and confidence (conf) are calculated. The symbolic sequence transforms the continuous sequence results into an ordered state, while the three-dimensional feature vector collectively describes the concentration level, rate of change, and trend direction from different perspectives. Through the statistical results of support (sup) and confidence (conf), the co-occurrence relationship between different feature states can be quantified. When significant deviations occur, it helps to indicate possible changes in the operating state.

[0018] 3. This application identifies anomalous candidate segments by using the results of support (sup) and confidence (conf), corrects these data, and then uses the corrected gas concentration data. The prediction model is input to obtain a trend vector T(t+Δt), and a hierarchical diagnosis is performed based on the prediction results. This allows anomalies to be located and corrected in the time series, and the data entering the model is more suitable in terms of continuity and stability. The prediction results are transformed into intuitive hierarchical outputs in a hierarchical form, making them easy to understand and use in operation and maintenance. The overall processing flow, from time series modeling and feature statistics to prediction and hierarchical classification, forms a progressively connected process, making the monitoring results more closely reflect the equipment's operating status. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart of a transformer oil chromatographic trend monitoring method based on multi-scale analysis provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application are described clearly and completely below. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of this application.

[0022] like Figure 1 As shown in the figure, this application provides a method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis, including the following steps: S1, collect time series data G(t) of gas concentration output by the online monitoring device for transformer oil chromatography; the characteristic gases corresponding to the time series data include hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4), carbon monoxide (CO), and carbon dioxide (CO2). S2, perform multi-scale sliding window piecewise fitting on the time series data G(t) of gas concentration to obtain piecewise fitted data; S3, perform clustering on the piecewise fitted data and convert the clustering results into a symbolic sequence Σ(t); S4. Based on the symbolic sequence Σ(t) and the concentration change rate R(t) and the second derivative of concentration A(t) calculated from G(t), a three-dimensional feature vector <G, R, A> composed of gas concentration, concentration change rate and second derivative of concentration is constructed, and the support sup and confidence conf are calculated. S5 identifies anomalous candidate segments based on the calculated support (sup) and confidence (conf), and corrects the anomalous candidate segments to obtain corrected gas concentration data. ; S6, for the corrected gas concentration data Perform trend modeling and prediction, output trend vector T(t+Δt), and perform hierarchical diagnosis based on the trend results.

[0023] The method in this embodiment is based on the time series data G(t) of gas concentration. It is no longer limited to single-point concentration values ​​or static ratios, but instead obtains piecewise fitted data through multi-scale sliding window piecewise fitting. The introduction of multiple scales helps to simultaneously reflect long-term trends and short-term fluctuations, while piecewise fitting divides the continuous sequence into several segments and expresses the changing characteristics of that time period by the overall trend within each segment. This not only reflects the dynamic characteristics of gas concentration over time but also facilitates comparison between different segments, providing a consistent input for subsequent processing.

[0024] Secondly, after obtaining the piecewise fitted data, this embodiment further employs clustering to generate a symbolic sequence Σ(t), and on this basis, constructs a three-dimensional feature vector <G, R, A> composed of gas concentration, concentration change rate, and second derivative of concentration. Subsequently, the support (sup) and confidence (conf) are calculated. The symbolic sequence transforms the continuous sequence results into an ordered state, while the three-dimensional feature vector collectively describes the concentration level, rate of change, and trend direction from different perspectives. Through the statistical results of support (sup) and confidence (conf), the co-occurrence relationship between different feature states can be quantified. When significant deviations occur, it helps to indicate possible changes in the operating state.

[0025] Finally, this embodiment identifies anomalous candidate segments based on the support (sup) and confidence (conf) results, corrects these data, and then uses the corrected gas concentration data. The prediction model is input to obtain a trend vector T(t+Δt), and a hierarchical diagnosis is performed based on the prediction results. This allows anomalies to be located and corrected in the time series, and the data entering the model is more suitable in terms of continuity and stability. The prediction results are transformed into intuitive hierarchical outputs in a hierarchical form, making them easy to understand and use in operation and maintenance. The overall processing flow, from time series modeling and feature statistics to prediction and hierarchical classification, forms a progressively connected process, making the monitoring results more closely reflect the equipment's operating status.

[0026] Furthermore, in some embodiments, the time series data G(t) of gas concentration is obtained by continuously collecting the concentrations of hydrogen, methane, acetylene, ethylene, carbon monoxide and carbon dioxide output by the online oil chromatography monitoring device at preset time intervals Δt, wherein Δt is 1 to 10 minutes.

[0027] In this embodiment, the time series data G(t) of gas concentration is obtained by continuously collecting the concentrations of hydrogen, methane, acetylene, ethylene, carbon monoxide, and carbon dioxide output from the online oil chromatography monitoring device at preset time intervals Δt (1–10 minutes). The fixed sampling interval provides a unified time reference for subsequent data processing, ensuring a consistent time scale when performing sliding window operations and change calculations, and reducing the interpolation requirements introduced by uneven intervals.

[0028] The value of Δt ranges from 1 to 10 minutes, which can cover the gas change rate of the transformer under common operating conditions and is compatible with the operating frequency of the online chromatography device. If the interval is too long, short-term fluctuations may be missed; if the interval is too short, noise and data volume may increase. Controlling Δt within this range helps to achieve a reasonable balance between data accuracy and system load.

[0029] Furthermore, the simultaneous recording of the six types of gases at the same time intervals ensures comparability of the relative changes between different gases, such as synchronous increases or sequential changes within the same time period. Since each gas has different indicative significance in fault types such as insulation aging, overheating, or discharge, this simultaneous recording method makes the time series available not only at the single-gas level but also at the multi-gas combination level, providing relatively consistent and comparable data input for subsequent trend modeling and diagnosis.

[0030] In some embodiments, the multi-scale sliding window piecewise fitting is implemented by including the following steps: S21, divide the time series data G(t) of gas concentration according to the short-time scale window length L1 and the long-time scale window length L2, and set the overlap rate r between adjacent windows; S22, within each window, perform fitting operations on the divided gas concentration data. The fitting operations include establishing a linear function or a polynomial function and extracting the corresponding parameter vector from the function. S23, use the set of parameter vectors corresponding to each window as piecewise fitting data.

[0031] In this embodiment, by setting a short-term window L1 and a long-term window L2 on the time series G(t) and introducing an overlap rate r between adjacent windows, the series is divided into several local segments. The short-term window is more sensitive to sudden fluctuations and can reflect rapid local changes; the long-term window is suitable for observing slow background trends. Using the two types of windows in combination helps to take into account different time characteristics and reduce the possibility of missing key signals due to a single scale. The overlap rate setting, to some extent, mitigates abrupt changes at the segment boundaries, making adjacent segments more consistent in numerical performance and facilitating cross-window comparisons.

[0032] Furthermore, within each window, a fitting operation is performed to transform scattered sampling points into a small number of parameter vectors to represent the overall trend of that segment. This approach reduces the interference of random fluctuations on trend representation while providing consistent feature dimensions for different windows, facilitating subsequent merging and analysis. After fitting and overlapping processing, the resulting segmented data not only retains temporal information but also includes a parameterized description of the segment's shape, making the data structurally more compact.

[0033] Therefore, the final piecewise fitted data has a clear temporal interval division, relatively simple feature representation, and relatively stable numerical performance. These characteristics provide suitable input conditions for subsequent clustering and trend analysis, and to some extent help improve the stability and reference value of subsequent analysis results.

[0034] In some embodiments, the short-timescale window length L1 ranges from 50 to 100 sampling points, the long-timescale window length L2 ranges from 300 to 500 sampling points, and the overlap rate r of adjacent windows is 20% to 40%.

[0035] In this embodiment, the short-term window length L1 is set between 50 and 100 sampling points, ensuring that the number of sampling points within the window meets the needs of the fitting operation without masking short-term fluctuations due to excessive samples. This setting helps capture rapid changes in gas concentration within a local time period. The long-term window L2 has a length of 300 to 500 sampling points, covering a longer operating range, making it easier to observe the overall evolution trend of concentration over time and mitigating the impact of occasional noise to some extent. The combined use of short-term and long-term windows allows the analysis results to simultaneously reflect rapid fluctuations and long-term trends, complementing each other and avoiding biases caused by a single scale.

[0036] Setting an overlap rate of 20% to 40% between adjacent windows makes the numerical representation between segments more continuous. Moderate overlap can mitigate abrupt changes at the segmentation boundaries, making comparisons between adjacent segments smoother. At the same time, maintaining this range avoids redundant computation problems caused by excessive overlap, keeping the computational load and data utilization at a suitable level.

[0037] The segmented data obtained through the above settings are continuous in time and relatively stable in numerical value, making them more suitable as input conditions for subsequent clustering and trend analysis. In application, this method can, to some extent, reduce noise amplification caused by excessively short windows or detail obscuration caused by excessively long windows, making the analysis results more valuable.

[0038] Specifically, step S3 includes the following steps: S31, use the segmented fitted data as clustering input; S52, the k-means clustering algorithm is used to classify the segmented fitted data; S53, determine the number of clusters k based on the maximization result of the silhouette coefficient Sc(k); S54. Assign symbolic identifiers to the cluster output categories and arrange them in chronological order to generate a symbolic sequence Σ(t).

[0039] The k-means clustering algorithm uses Euclidean distance as a measure of similarity between samples. It takes the parameter vector of the piecewise fitted data as input and gradually converges to obtain the clustering result through iterative calculation. Since the piecewise fitted data has relatively fixed dimensions, k-means clustering can classify segments with similar shapes while ensuring computational efficiency.

[0040] The silhouette coefficient Sc(k) ranges from -1 to 1 and is used to measure the compactness and separation of clustering results. When Sc(k) is close to 1, it indicates that the similarity of samples within a cluster is high and they are clearly distinguishable from other categories; when Sc(k) is close to 0, it indicates that the distinction between different categories is not obvious; when Sc(k) is negative, it indicates that there may be misassignment in the clustering. In this embodiment, Sc(k) is calculated under different numbers of clusters, and the number of clusters corresponding to the maximum value is selected as the final result to obtain a better category division.

[0041] Furthermore, after clustering, the system assigns a symbol to each category and then arranges these symbols sequentially according to the chronological order of the data segments to generate a symbolic sequence Σ(t). This symbolic sequence not only preserves the chronological order information but also transforms the originally continuous numerical data into a finite-state sequence, facilitating subsequent sequence-based statistical analysis and pattern recognition.

[0042] In this embodiment, segmented fitted data is used as input for clustering, enabling comparison of different window segments on a uniform numerical scale. Using the k-means clustering method, similar local trends can be grouped into the same category with lower computational overhead, thus reducing a large number of continuous segments into a limited number of classes.

[0043] Subsequently, the results of different cluster numbers are evaluated using the silhouette coefficient Sc(k), and an appropriate cluster number k is selected. This achieves a balance between intra-class similarity and inter-class differences, providing a more objective basis compared to simply relying on empirical settings.

[0044] After clustering, the output categories are assigned corresponding symbols and arranged in chronological order to obtain a symbolic sequence Σ(t). This method transforms a continuous curve into a finite symbolic sequence, preserving the temporal order while forming a statistically representative state representation. After symbolization, the sequence can be directly used to analyze features such as state duration and transition locations, thereby reducing the interference of random fluctuations on subsequent pattern recognition to some extent.

[0045] In some embodiments, the construction of the three-dimensional feature vectors <G, R, A> in step S4 includes: The time series data of gas concentration G(t) at sampling time t i The gas concentration value G(t) i ) as the G component in the three-dimensional feature vector <G, R, A>; The rate of concentration change R(t) is calculated according to the following formula. i ): ; Wherein, R(t) i ) indicates at time t i The rate of concentration change; The second derivative of the concentration, A(t), is calculated using the following formula. i ): ; Where A(t) i ) indicates at time t i The rate of change of the concentration; G(t) i ), R(t) i ) and A(t i The three-dimensional feature vectors <G, R, A> are combined and mapped to the corresponding segments of the symbolic sequence Σ(t).

[0046] In this embodiment, by constructing a three-dimensional feature vector <G, R, A>, the concentration level, the rate of concentration change, and the trend of rate change can be uniformly represented under the same coordinate system. This allows for the numerical reflection of different change patterns, such as slow increases, rapid increases, or accelerated changes.

[0047] R(t i ) and A(t i The calculation of R(t) is based on adjacent sampling points, making the operation relatively simple and less susceptible to long-term drift or overall scale differences. Under the condition of uniform sampling intervals, R(t) i ) and A(t i Maintaining consistency in the expression of data makes comparisons between different gases or different time periods more direct, thereby reducing the need for additional normalization processing to some extent and simplifying data preprocessing.

[0048] After obtaining <G, R, A>, they are aggregated according to the segmented results of the symbolic sequence Σ(t), which can be used to analyze the characteristic performance and changes in each state segment. For example, when R is large and A is positive in a certain segment, it can be judged that the segment is in an upward trend; while if R decreases or A changes from positive to negative in subsequent segments, it indicates that the trend direction may have reversed. In this way, the time series is transformed into state segments that contain both numerical features with physical meaning and retain the temporal order, providing more intuitive input conditions for subsequent statistics and discrimination.

[0049] After constructing the three-dimensional feature vectors <G, R, A>, the above embodiment establishes a correspondence between each feature vector and the segmented result of the symbolic sequence Σ(t). Specifically, the symbolic sequence Σ(t) has divided the time series into several segments, each segment corresponding to a symbol identifier. All three-dimensional feature vectors generated within this segment are assigned this symbol identifier, thereby realizing the mapping between feature vectors and the symbolic sequence. In this way, the original level information of gas concentration (G) is preserved, while dynamic change features (R and A) are introduced and combined with the symbolic sequence Σ(t) to form signed multidimensional feature data, which can be used for subsequent calculation of support (sup) and confidence (conf).

[0050] In some embodiments, the calculation methods for support (sup) and confidence (conf) in step S4 include: The three-dimensional feature vectors <G, R, A> are discretized according to the segmentation results of the symbolic sequence Σ(t) to obtain the itemset sequence; In an itemset sequence, count the number of times any itemset X and itemset Y appear simultaneously, count(X∪Y), and count the number of times itemset X appears, count(X). Calculate the support and confidence scores according to the association rule formula: ; ; Where N represents the total number of items in the itemset sequence, X and Y represent itemsets in the symbolic sequence Σ(t), count(X∪Y) represents the number of times itemset X and itemset Y appear simultaneously, and count(X) represents the number of times itemset X appears.

[0051] In the above embodiments, support Represents a set rule The frequency of occurrence in all data, i.e., the proportion of itemsets containing both X and Y in the entire set, is used to measure the generality of the rule; confidence level. represents the probability that itemset Y will occur given itemset X, and is used to measure the reliability of the rule. Support (sup) and confidence (conf) are defined according to standard calculation formulas for association rule mining, and are used to describe the association between three-dimensional feature vectors. By calculating sup and conf, the association between gas concentration patterns can be measured, providing a basis for subsequent anomaly detection.

[0052] In this embodiment, the <G, R, A> eigenvectors are divided into discrete categories according to the segmentation results of the symbolic sequence Σ(t), thus forming a finite itemset. In this way, originally continuous values ​​are grouped into fixed intervals, avoiding the amplification effect caused by small measurement fluctuations, making subsequent statistical analysis more stable at the numerical level. The establishment of itemsets allows comparisons between different time segments to be based on a unified category system, and also facilitates the comparison and analysis between different gases.

[0053] Based on this, support (sup) and confidence (conf) are calculated according to the frequency of occurrence of itemsets and the frequency of joint occurrences. Support reflects the proportion of a particular itemset in the overall data, while confidence describes the probability of another item occurring simultaneously under given conditions. These two metrics allow us to express the prevalence and correlation of different combinations of states or features during the process using statistical methods, ensuring that the analysis results are based on the distribution characteristics of the data, rather than relying on empirical thresholds.

[0054] When operating conditions change, combinations that were previously frequent may decrease, while some less frequent combinations may increase. This change can be reflected in the numerical differences between sup and conf, revealing potential differences in operating trends even if the absolute values ​​of the concentrations remain within the normal range. Thus, the combinational relationships of <G, R, A> are transformed into a recordable and comparable form through itemsets and statistical indicators, providing a relatively objective reference for the monitoring process.

[0055] Furthermore, in step S5, the method for determining abnormal candidate segments includes: The calculated support score (sup) and confidence score (conf) are compared with the baseline support score (sup0) and baseline confidence score (conf0) obtained based on historical normal operation data. A segment is considered an abnormal candidate segment when one of the following conditions is met: or ; Where, sup0 represents the baseline support value obtained based on historical normal operation data, conf0 represents the baseline confidence value obtained based on historical normal operation data, α represents the bias judgment threshold of support, which ranges from 0.05 to 0.20, and β represents the bias judgment threshold of confidence, which ranges from 0.05 to 0.20.

[0056] In this embodiment, the determination of abnormal candidate segments does not rely on a fixed threshold. Instead, it compares the support (sup) and confidence (conf) calculated segment by segment with the baseline values ​​(sup0 and sup0) obtained from statistics of the same device during normal operation. The determination condition adopts the absolute deviation form, i.e., |sup-sup0|≥α or |conf-conf0|≥β, where the values ​​of α and β are set to the range of 0.05 to 0.20. In this way, the determination basis comes from the historical statistical patterns of the device itself, which is more in line with the operating background and individual differences.

[0057] This method of judgment allows normal fluctuations under different operating conditions to be included in the baseline range, thereby reducing false alarms caused by differences in operating conditions to some extent. Support reflects the proportion of a combination in the overall data, while confidence reflects the probability of another combination occurring simultaneously under specific conditions. When used together, they can more intuitively reveal changes in the combination relationships. For example, when combinations that were previously frequent decrease in number, or combinations that were previously infrequent increase in number, this shift will be reflected in the changes in the values ​​of support and confidence, thus indicating potential operational anomalies.

[0058] Furthermore, introducing absolute deviation judgment and setting threshold intervals can statistically filter out small fluctuations, making candidate segment identification more focused on changes with practical significance. This approach transforms anomaly detection into deviation comparison based on historical norms, which helps improve the reliability of the identification results and provides relatively robust input conditions for subsequent data correction and trend analysis.

[0059] Furthermore, in step S5, the data correction method is as follows: Gas concentration data identified as abnormal candidate segments are filtered using Kalman filtering to reduce the impact of random noise on the data. After filtering, regression compensation is performed on the residual values ​​of the outlier candidate segments to ensure the corrected gas concentration data. The data trend remains continuous with that of adjacent normal segments; When the corrected gas concentration data Reference data G of adjacent normal segments ref (t i The mean square error (MSE) of a given condition is satisfied when: ; The correction is deemed valid; among which, G ref (t i ) represents the reference sequence, n represents the number of sample points involved in the calculation, and γ represents the preset error threshold, which is set to 10. -3 ~10 -2 .

[0060] In this embodiment, the correction operation is only performed on data segments identified as anomalous candidate segments. Segments not identified as anomalous remain in their original state to avoid deviations in normal data due to additional processing. Using Kalman filtering on candidate segments can, to some extent, reduce the impact of random noise and isolated outliers, while maintaining the continuity and local variation characteristics within the segment, making the data sequence smoother over time.

[0061] After filtering, the corrected result is compared with the reference sequence of the adjacent normal segment, and the mean squared error (MSE) is used as the criterion to determine whether it is less than a preset threshold γ. This ensures that the corrected result is consistent with the reference state of the equipment under the same operating background, while utilizing the sensitivity of the MSE to deviations to avoid over-smoothing or under-correction. The threshold γ, combined with the setting of the sample size n, provides a quantitative basis for determining whether the corrected result is acceptable.

[0062] After the above processing, the data used for subsequent modeling is improved in terms of fluctuation amplitude and continuity, making it easier to compare the numerical performance of outlier and normal segments. This approach reduces spurious trends caused by noise and isolated outliers, and to some extent improves the stability of subsequent trend modeling and analysis.

[0063] In some embodiments, step S6 specifically includes: For the corrected gas concentration data A time series forecasting model is used for modeling and forecasting to obtain the trend vector T(t+Δt); the time series forecasting model adopts the autoregressive moving average model ARIMA. The gas concentration growth rate r is calculated based on the trend vector T(t+Δt). g ; gas concentration growth rate r g The unit is % / day; When r g When the risk level is less than the threshold α1, it is classified as a Level 1 risk with corresponding low risk. When r g When the risk level is between α1 and β1, it is classified as a secondary risk within the medium-risk category. When r g When β1 is greater than or equal to β1, it is classified as a level 3 risk corresponding to high risk; Among them, the value of α1 ranges from 0.5% to 1.0% per day, and the value of β1 ranges from 1.0% to 2.0% per day.

[0064] The Autoregressive Moving Average (ARIMA) model is denoted as ARIMA(p,d,q), where p, d, and q represent the autoregressive order, differencing order, and moving average order, respectively. The modeling process includes: first, processing the corrected gas concentration sequence... Perform a stationarity test and, if necessary, perform differencing; then, combine the autocorrelation function and partial autocorrelation function to determine the range of p and q, and use information criteria (such as AIC or BIC) to select the optimal model; finally, perform parameter estimation and fitting on the data, and output the predicted value T(t+Δt) for future time based on the fitted model.

[0065] In this embodiment, the corrected time series The data is input into a time series forecasting model, such as using the ARIMA method, to obtain the trend vector T(t+Δt) for future times. This modeling approach utilizes the time dependencies of the series itself for calculation, and compared to single-point observations, it can better reflect the continuity of concentration over time, thus making the forecast results more meaningful for short-term trends.

[0066] After obtaining the prediction results, the trend vector is transformed into the gas concentration growth rate r. g The data is expressed as a percentage per day. Through rate-based processing, trend changes in different gases or from different devices can be compared on a uniform scale. Using a rate expression with the same time step also facilitates comparison with historical data, providing a more intuitive reflection of the rate of concentration change.

[0067] Furthermore, in this embodiment, the growth rate r is affected by setting thresholds α1 and β1. g Perform grading: when r g A risk level below α1 is classified as Level 1 risk, a risk level between α1 and α1 as Level 2 risk, and a risk level above or equal to β1 as Level 3 risk. This tiered risk assessment method, compared to a single threshold assessment, is more effective at mitigating the impact of short-term fluctuations, resulting in more stable numerical values. In this way, the predicted results are transformed into specific risk levels, providing an intuitive reference for operational monitoring and facilitating comparisons across multiple devices.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis, characterized in that, Includes the following steps: S1, Collect time series data G(t) of gas concentration output by the online transformer oil chromatography monitoring device; wherein the characteristic gases corresponding to the time series data include hydrogen, methane, acetylene, ethylene, carbon monoxide and carbon dioxide; S2, perform multi-scale sliding window piecewise fitting on the time series data G(t) of the gas concentration to obtain piecewise fitted data; S3, perform clustering processing on the segmented fitting data, and convert the clustering results into a symbolic sequence Σ(t); S4. Based on the symbolic sequence Σ(t) and the concentration change rate R(t) and the second derivative of concentration A(t) calculated from G(t), construct a three-dimensional feature vector <G, R, A> composed of gas concentration, concentration change rate and second derivative of concentration, and calculate the support sup and confidence conf. S5. Based on the calculation results of the support (sup) and confidence (conf), identify anomalous candidate segments, and perform data correction on the anomalous candidate segments to obtain corrected gas concentration data. ; S6, the corrected gas concentration data Perform trend modeling and prediction, output trend vector T(t+Δt), and perform hierarchical diagnosis based on trend results; The implementation method of the multi-scale sliding window piecewise fitting includes the following steps: S21, divide the time series data G(t) of gas concentration according to the short-time scale window length L1 and the long-time scale window length L2, and set the overlap rate r between adjacent windows; S22, within each window, a fitting operation is performed on the divided gas concentration data. The fitting operation includes establishing a linear function or a polynomial function and extracting the corresponding parameter vector from the function. S23, use the set of parameter vectors corresponding to each window as piecewise fitting data; Step S3 includes the following steps: S31, use the segmented fitting data as clustering input; S32, The k-means clustering algorithm is used to classify the segmented fitted data; S33, determine the number of clusters k based on the maximization result of the silhouette coefficient Sc(k); S34. Assign symbolic labels to the cluster output categories and arrange them in chronological order to generate a symbolic sequence Σ(t); The methods for constructing the three-dimensional feature vectors <G, R, A> in step S4 include: The time series data of gas concentration G(t) at sampling time t i The gas concentration value G(t) i ) as the G component in the three-dimensional feature vector <G, R, A>; The rate of concentration change R(t) is calculated according to the following formula. i ): ； Wherein, R(t) i ) indicates at time t i The rate of concentration change; The second derivative of the concentration, A(t), is calculated using the following formula. i ): ； Where A(t) i ) indicates at time t i The rate of change of the concentration; G(t) i ), R(t) i ) and A(t i The three-dimensional feature vectors <G, R, A> are combined and mapped to the corresponding segments of the symbolic sequence Σ(t).

2. The method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis according to claim 1, characterized in that, The time series data G(t) of the gas concentration is obtained by continuously collecting the concentrations of hydrogen, methane, acetylene, ethylene, carbon monoxide and carbon dioxide output by the online oil chromatography monitoring device at preset time intervals Δt, where Δt is 1 to 10 minutes.

3. The method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis according to claim 1, characterized in that, The short-term window length L1 ranges from 50 to 100 sampling points, the long-term window length L2 ranges from 300 to 500 sampling points, and the overlap rate r of adjacent windows is 20% to 40%.

4. The method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis according to claim 1, characterized in that, The calculation methods for support (sup) and confidence (conf) in step S4 include: The three-dimensional feature vectors <G, R, A> are discretized according to the segmentation result of the symbolic sequence Σ(t) to obtain the itemset sequence; In the itemset sequence, count the number of times any itemset X and itemset Y appear simultaneously, count(X∪Y), and count the number of times itemset X appears, count(X). Calculate the support and confidence scores according to the association rule formula: ； ； Where N represents the total number of items in the itemset sequence, X and Y represent itemsets in the symbolic sequence Σ(t), count(X∪Y) represents the number of times itemset X and itemset Y appear simultaneously, and count(X) represents the number of times itemset X appears.

5. The method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis according to claim 4, characterized in that, In step S5, the methods for determining abnormal candidate segments include: The calculated support score (sup) and confidence score (conf) are compared with the baseline support score (sup0) and baseline confidence score (conf0) obtained based on historical normal operation data. A segment is considered an abnormal candidate segment when one of the following conditions is met: or ; Where, sup0 represents the baseline support value obtained based on historical normal operation data, conf0 represents the baseline confidence value obtained based on historical normal operation data, α represents the bias judgment threshold of support, which ranges from 0.05 to 0.20, and β represents the bias judgment threshold of confidence, which ranges from 0.05 to 0.

20.

6. The method for monitoring the chromatographic trend of transformer oil based on multi-scale analysis according to claim 1, characterized in that, In step S5, the data correction method is as follows: The gas concentration data identified as abnormal candidate segments are filtered using Kalman filtering or moving average filtering. After filtering, regression compensation is performed on the residual values ​​of the abnormal candidate segments to ensure the corrected gas concentration data. The data trend remains continuous with that of adjacent normal segments; When the corrected gas concentration data Reference data G of adjacent normal segments ref (t i The mean square error (MSE) of a given condition is satisfied when: ； The correction is deemed valid; wherein, G ref (t i ) represents the reference sequence, n represents the number of sample points involved in the calculation, and γ represents the preset error threshold, which is set to 10. -3 ~10 -2 .

7. The method for monitoring chromatographic trends of transformer oil based on multi-scale analysis according to claim 1, characterized in that, Step S6 is as follows: For the corrected gas concentration data A time series forecasting model is used for modeling and forecasting to obtain the trend vector T(t+Δt); the time series forecasting model adopts the autoregressive moving average model ARIMA. The gas concentration growth rate r is calculated based on the trend vector T(t+Δt). g ; gas concentration growth rate r g The unit is % / day; When r g When the risk level is less than the threshold α1, it is classified as a Level 1 risk with corresponding low risk. When r g When the risk level is between α1 and β1, it is classified as a secondary risk within the medium-risk category. When r g When β1 is greater than or equal to β1, it is classified as a level 3 risk corresponding to high risk; Among them, the value of α1 ranges from 0.5% to 1.0% per day, and the value of β1 ranges from 1.0% to 2.0% per day.

Images