An adaptive detection method for electrical steel strip transformers

Abstract

This invention discloses an adaptive detection method for electrical steel strip transformers. In the field of transformer condition monitoring, the method synchronously acquires input power, output power, load current, and winding resistance through time-series alignment, constructing a correlation mapping between the total active power loss sequence and the winding loss sequence. Based on historical normal operation data, detrending and scaling normalization are performed on the two sequences to eliminate baseline drift and unify dimensions. Furthermore, a sliding window is used to extract local feature points, and discrete trajectory curves are plotted in a two-dimensional coordinate system composed of total active power loss and winding loss. Polar coordinate sector intervals are used to obtain the radial width distribution of each angular sector, and permutation entropy is introduced to quantify the complexity of the trajectory shape, which serves as a loss loop anomaly index for real-time locking and alarm output of abnormal states. The permutation entropy trend of continuous windows is tracked, and a performance degradation warning is generated when the cumulative increase exceeds a preset threshold, taking into account both instantaneous fault detection and gradual aging assessment.

Description

Technical Field

[0001] This invention belongs to the field of transformer condition detection technology, specifically, it relates to an adaptive detection method for electrical steel strip transformers. Background Technology

[0002] Data-driven adaptive detection technology has gradually become a research hotspot. By deeply integrating time-series signal processing and intelligent analysis methods, it can achieve dynamic characterization of transformer loss characteristics and accurate identification of operating status, providing important technical support for proactive operation and maintenance and full life cycle management of power equipment.

[0003] In existing electrical steel strip transformer detection technologies, traditional anomaly monitoring methods typically rely on pre-set static thresholds or simple difference comparisons, such as directly monitoring whether the total active power loss exceeds a certain fixed upper limit, or analyzing changes in load current and winding resistance separately. Traditional methods neglect the strict temporal alignment between input and output power, easily leading to loss calculation errors due to sampling delays or data asynchrony. Secondly, existing technologies rarely consider the daily cycle or seasonal trend components inherent in the loss data. Factors such as grid load patterns can cause systematic shifts in losses; directly using raw loss values ​​for judgment can generate numerous false alarms due to trend interference, especially in transformers with variable current characteristics. When a transformer is under light load or no load, static thresholds can easily misjudge normal fluctuations as abnormalities. In addition, traditional methods usually only focus on the absolute value of losses in a single dimension, and fail to conduct joint analysis of total active power loss and winding loss in phase space, thus failing to capture the dynamic coupling relationship between the two. In fact, core loss and winding loss will show drastically different relative change characteristics under different fault modes, and it is difficult to distinguish them based on their respective amplitudes alone. Existing anomaly detection indicators are mostly instantaneous or short-term average values, making it impossible for the system to perceive anomalies in the loss trajectory. They lack adaptive normalization and detrending mechanisms, and thresholds must be recalibrated for transformers of different specifications and aging levels, resulting in poor field adaptability.

[0004] To address the aforementioned problems, this invention proposes an adaptive detection method for electrical steel strip transformers. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an adaptive detection method for electrical steel strip transformers, solving the problems of existing technologies being unable to dynamically capture the coupling characteristics between losses, being susceptible to trend interference, and being unable to quantify abnormal loss loop morphology.

[0006] The objective of this invention can be achieved through the following technical solutions: An adaptive detection method for electrical steel strip transformers, the method comprising: Step 1: Lock the input power and output power in a time-aligned manner to generate the total active power loss sequence H1 associated with the time sequence, and simultaneously lock the load current and winding resistance in real time to generate the winding loss sequence H2 corresponding to the total active power loss sequence. Step 2: Based on the historical average values ​​of H1 and H2 at the corresponding time, detrending is performed on each of them. Normalization is then performed using the scale parameters of the two sequences after detrending to generate the normalized total active power loss sequence G1 and the normalized winding loss sequence G2. Step 3: Lock the mean values ​​of sliding windows G1 and G2 within the sliding window respectively. Use the different mean values ​​of sliding windows and the data contained in G1 and G2 within different sliding windows to perform point pair extraction, construct a set of point pairs, and plot the discrete trajectory curve in the pre-constructed two-dimensional coordinate system. Step four: Based on the discrete trajectory curve, perform angular division to determine several sector intervals, determine the maximum and minimum extreme radii based on the extreme radii boundaries and angular step size of the sector intervals, further evaluate the loss loop anomaly index of the corresponding sliding window, lock the abnormal state in real time, and generate an abnormal alarm signal.

[0007] As a further aspect of the present invention, the specific method for generating the time-series-related total active power loss sequence H1 in step one is as follows: Taking the current time as the start time of detection, denoted as t1, the input power and output power of the electrical steel strip transformer are collected at a preset sampling frequency to obtain the input power sequence P_in1, P_in2, ..., P_inj and the output power sequence P_out1, P_out2, ..., P_outj, where j is the total number of samples, which increases with time. Calculate the difference between input power and output power at each sampling time to determine the total active power loss sequence H1=h1,h2,...,hj.

[0008] As a further aspect of the present invention, the specific method for generating the winding loss sequence H2 corresponding to the total active power loss sequence in step one is as follows: Determine the load current sequence D1, D2, ..., Dj and winding resistance R corresponding to the input power sequence and output power sequence timing. The winding resistance R is taken as the average resistance value of the electrical steel strip transformer under normal operating conditions and is considered as a known value. Using pi = v × (Di) 2 ×R calculates the winding loss at each sampling time, forming a winding loss sequence H2=p1,p2,...,pj, where v is the total number of phase windings of the electrical steel strip transformer, i is the counting index, and 1≤i≤j.

[0009] As a further aspect of the present invention, in step two, the specific method for performing detrending based on the historical averages of H1 and H2 at corresponding times is as follows: The total active power loss sequence H1 = h1, h2, ..., hj is analyzed. Similarly, the winding loss sequence H2 = p1, p2, ..., pj is obtained; For any sampling time i, lock its corresponding time within the day, and perform an average of the total active power loss of the electrical steel strip transformer at the same time within the day under the historical normal operation state to obtain the average total active power loss μh_i, where the averaged total active power loss is the n closest to the current time, and n is a preset integer. Similarly, determine the mean winding loss μp_i; Detrending is performed on the total active power loss sequence H1 and the winding loss sequence H2 using hi'=hi-μh_i and pi'=pi-μp_i respectively, to obtain the detrended total active power loss H1'=h1',h2',...,hj' and the detrended winding loss sequence H2'=p1',p2',...,pj'.

[0010] As a further aspect of the present invention, the specific method for performing normalization using the scale parameters of the two sequences after detrending in step two is as follows: Obtain the scale parameter of the detrended total active power loss H1', ​​denoted as the first scale parameter Ch1, and obtain the scale parameter of the detrended winding loss sequence H2', denoted as the second scale parameter Ch2. The first scale parameter Ch1 is at least one of the standard deviation, mean absolute deviation, or interquartile range of the detrended total active power loss H1', ​​and the second scale parameter Ch2 is similar. Normalization of H1' and H2' is performed using gi=hi' / Ch1 and qi=pi' / Ch2 respectively, resulting in the normalized total active power loss sequence G1=g1,g2,...,gj and the normalized winding loss sequence G2=q1,q2,...,qj.

[0011] As a further aspect of the present invention, in step three, point pair extraction is performed using different sliding window averages and the data contained in G1 and G2 within different sliding windows. The specific method for constructing the point pair set is as follows: Extract the sliding window W and the sliding step size Δ, where W and Δ are both preset positive integers, and Δ≤W; Starting from the detection start time t1, the sliding windows are sequentially slid with a sliding step size Δ to obtain K sliding windows, where, Indicates rounding down; For any k-th sliding window, take the sampling time range [t_min, t_max] it covers, and determine the normalized total active power loss corresponding to [t_min, t_max] from the normalized total active power loss sequence G1. Calculate the mean, denoted as the sliding window mean μG1_k, k=1,2,...,K; Similarly, the normalized winding loss sequence G2 is processed to determine the sliding window mean μG2_k; Take the normalized total active power loss gi and normalized winding loss qi corresponding to any sampling time i within the sampling time range [t_min, t_max]. The x-coordinate Xi and y-coordinate Yi are constructed using Xi=gi-μG1_k and Yi=qi-μG2_k respectively, generating a pair of points (Xi,Yi); Similarly, synchronization processing is performed on all sampling times [t_min, t_max] within the k-th sliding window to obtain the point pair set DH_k; Similarly, processing the K sliding windows yields a set of K point pairs: DH_1, DH_2, ..., DH_K.

[0012] As a further aspect of the present invention, the specific method for plotting the discrete trajectory curve in the pre-constructed two-dimensional coordinate system in step three is as follows: Extract the set of point pairs DH_k associated with any k-th sliding window; A two-dimensional coordinate system is constructed with the local deviation of total active power loss as the horizontal axis and the local deviation of winding loss as the vertical axis. Get any pair of points (Xi, Yi) from the set of point pairs DH_k, align them with the horizontal and vertical axes of the two-dimensional coordinate system, and plot the data points; Similarly, based on the set of point pairs DH_k, all data points are plotted in a two-dimensional coordinate system, and all data points are connected in the order of sampling time to obtain the discrete trajectory curve, denoted as LS_k.

[0013] As a further aspect of the present invention, the specific method for determining the maximum and minimum polar radii based on the polar radius boundary and angular step size of the sector interval in step four is as follows: With the origin of the two-dimensional coordinate system as the pole, the two-dimensional coordinate system is divided into M sector intervals according to a preset angular step Δθ, where M=2π / Δθ, Δθ∈[1°,5°]; For each sector interval, traverse all data points on the discrete trajectory curve LS_k, filter out the data points where the polar corners are located in the corresponding sector interval, calculate the polar radius of each data point, and record the maximum polar radius R_max and the minimum polar radius R_min in the corresponding sector interval. If there are no data points in a certain sector interval, then the linear interpolation of the maximum extreme diameter R_max and the minimum extreme diameter R_min of the adjacent sector interval is taken as the estimated maximum and minimum extreme diameters of that sector interval.

[0014] As a further aspect of the present invention, in step four, the specific method for further evaluating the loss loop anomaly index corresponding to the sliding window, locking the abnormal state in real time, and generating an abnormal alarm signal is as follows: For any k-th sliding window, extract the difference between the maximum and minimum extreme radii of each sector interval, denoted as the radial width, and arrange them in angular order to form the radial width sequence Z1, Z2, ..., ZM; The radial width sequence Z1, Z2, ..., ZM is reconstructed in phase space. Based on the preset embedding dimension d and time delay τ, the radial width sequence is transformed into several radial width subsequences. The permutation patterns of numerical values ​​in each radial width subsequence are statistically analyzed, and the probability of each permutation pattern is calculated to obtain the permutation entropy ET_k of the k-th sliding window, which is used as the loss ring anomaly index of the k-th sliding window. Based on historical normal data of electrical steel strip transformers of the same specification and model under normal operating conditions, the baseline range of the permutation entropy is locked, and the permutation entropy distribution is calculated to obtain the lower limit threshold ET_min and the upper limit threshold ET_max. If the current sliding window's loss loop anomaly index ET_k < ET_min or ET_k > ET_max, then the current sliding window is determined to be abnormal, and an anomaly alarm signal is output.

[0015] As a further aspect of the present invention, step four also includes simultaneously performing trend analysis on the arrangement entropy of the continuous sliding windows. If the arrangement entropy continuously increases for more than the duration of a preset r sliding windows and the cumulative increase exceeds a preset percentage α, a performance degradation warning is generated. When an anomaly or performance degradation warning is triggered, an anomaly alarm signal is output.

[0016] The beneficial effects of this invention are: This invention decomposes the total active power loss and winding loss sequence through time alignment and synchronization locking, eliminating the interference of power fluctuations on loss assessment and thus improving the reliability of detection. It employs detrending and normalization processing to filter out deviations caused by historical averages, making loss data under different operating conditions comparable and enhancing adaptability. Furthermore, based on a sliding window to extract point pairs and construct discrete trajectory curves, the innovative mechanism of evaluating the loss loop anomaly index through geometric analysis of sector intervals eliminates the need for manually setting static thresholds, dynamically capturing minute changes in loss characteristics, and identifying abnormal states and generating alarm signals in real time and with high sensitivity. This achieves a fully adaptive detection process, avoiding the dependence on empirical parameters in traditional methods. The combination of trajectory curves and sector intervals makes anomaly quantification more intuitive. This invention introduces a detrending operation based on historical averages to remove the trend components caused by factors such as periodic load and ambient temperature from the total active power loss and winding loss of transformers. This makes the detrended loss more purely reflect the abnormal or degradation characteristics of the equipment itself. If normalization is performed using scale parameters such as standard deviation or interquartile range, loss sequences with different dimensions or fluctuation amplitudes are transformed to the same comparable scale, eliminating the magnitude differences between different characteristic quantities, enhancing the robustness of subsequent analysis, avoiding the masking of weak fault signals by normal fluctuations or leading to misjudgment, and improving the consistency of cross-time and cross-equipment comparisons. This invention uses a sliding window averaging process to remove DC bias or long-term trend terms from the normalized total active power loss sequence and winding loss sequence, enabling the extracted point pairs to truly reflect local fluctuation characteristics. This improves the sensitivity to fault characteristics or changes in operating status. By using point pair sets constructed within different windows, the curves are plotted in a two-dimensional coordinate system and connected according to the sampling time to form discrete trajectory curves. This intuitively reveals the dynamic correlation between total active power loss and winding loss. Abnormal patterns can be quickly identified through geometric features such as trajectory shape and direction. At the same time, the adjustability of the sliding window and step size gives it good adaptability, which is convenient for online monitoring and early warning. This invention improves the robustness and completeness of loss loop morphology feature extraction by dividing a two-dimensional coordinate system into sector intervals and adaptively determining the polar radius boundary, combined with linear interpolation to process data-free regions. On this basis, it uses permutation entropy as an anomaly index, which can sensitively capture the complex dynamic changes of radial width sequences without relying on specific distribution assumptions, effectively distinguishing normal fluctuations from real anomalies. Furthermore, it establishes a permutation entropy baseline threshold by combining historical normal data and introduces trend analysis of continuous sliding windows to lock in instantaneous abnormal states in real time and provide early warning of the continuous upward trend of permutation entropy, thereby achieving early identification of performance degradation. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the method described in this invention; Figure 2 This is a flowchart illustrating the method described in Embodiment 4 of the present invention; Figure 3 This is a flowchart illustrating the method described in Embodiment 5 of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this application provides an adaptive detection method for electrical steel strip transformers; As an embodiment 1 of this application, it specifically includes: Step 1: Lock the input power and output power in a time-aligned manner to generate the total active power loss sequence H1 associated with the time sequence, and simultaneously lock the load current and winding resistance in real time to generate the winding loss sequence H2 corresponding to the total active power loss sequence. Step 2: Based on the historical average values ​​of H1 and H2 at the corresponding time, detrending is performed on each of them. Normalization is then performed using the scale parameters of the two sequences after detrending to generate the normalized total active power loss sequence G1 and the normalized winding loss sequence G2. Step 3: Lock the mean values ​​of sliding windows G1 and G2 within the sliding window respectively. Use the different mean values ​​of sliding windows and the data contained in G1 and G2 within different sliding windows to perform point pair extraction, construct a set of point pairs, and plot the discrete trajectory curve in the pre-constructed two-dimensional coordinate system. Step four: Based on the discrete trajectory curve, perform angular division to determine several sector intervals, determine the maximum and minimum extreme radii based on the extreme radii boundaries and angular step size of the sector intervals, further evaluate the loss loop anomaly index of the corresponding sliding window, lock the abnormal state in real time, and generate an abnormal alarm signal.

[0021] Example 2

[0022] This embodiment, based on Embodiment 1, further discloses a time-series monitoring method for total active power loss and winding loss of a transformer, specifically including the following: First, identify the electrical steel strip transformer to be monitored, and take the current time as the detection start time t1, and perform the following example processing; First, the sampling frequency is set to once per minute. Then, based on this sampling frequency, the input power and output power of the electrical steel strip transformer are collected. The collected input power and output power are sorted according to the time sequence (sampling sequence) to obtain the input power sequence P_in1, P_in2, ..., P_inj and the output power sequence P_out1, P_out2, ..., P_outj, where j is the total number of samples, which increases over time. The total active power loss of the electrical steel strip transformer includes winding copper loss, core loss (eddy current + hysteresis), and stray loss. According to the law of conservation of energy, the input power minus the output power is the active power consumed by the equipment during operation, which can directly reflect the instantaneous value of the total loss. Therefore, the difference between the input power and the output power at each sampling time is calculated to determine the total active power loss sequence H1 = h1, h2, ..., hj.

[0023] Next, the load current sequences D1, D2, ..., Dj and winding resistance R corresponding to the input power sequence and output power sequence are acquired in real time according to the above sampling frequency. The load current is obtained by synchronously collecting the phase current of each phase on the secondary side of the electrical steel strip transformer using a current transformer and then converting it from analog to digital.

[0024] It should be noted that the winding resistance R is taken as the average resistance value of the electrical steel strip transformer under normal operating conditions, and is considered as a known value. Generally, the rated operating temperature of the electrical steel strip transformer is 75 degrees Celsius. If the actual operating temperature of the electrical steel strip transformer exceeds 75 degrees Celsius, the actual winding resistance R is estimated online based on the winding surface temperature measured by the pre-installed temperature sensor and the resistance-current iteration method.

[0025] Ultimately, by adopting pi = v × (Di) 2 ×R calculates the winding loss at any sampling time, forming a winding loss sequence H2=p1,p2,...,pj, where v is the total number of phase windings of the electrical steel strip transformer. For example, if it is a 3-phase electrical steel strip transformer, then v=3, and i is the counting index, 1≤i≤j.

[0026] Example 3

[0027] This embodiment further discloses a detrending and normalization preprocessing method based on historical mean, building upon embodiment 2. Specifically, it includes the following: Based on the content described in Example 2, the total active power loss sequence H1=h1,h2,...,hj and the winding loss sequence H2=p1,p2,...,pj are extracted; For any sampling time i in the total active power loss sequence H1, find the corresponding time (time point, such as 5:20 am) within a day for that sampling time i. Then, obtain the total active power loss of the electrical steel strip transformer at 5:20 am under the historical normal operation state from the cloud database, and perform an average operation, which is recorded as the average total active power loss μh_i. It should be noted that in the above operation, the total number of active power loss data for the average operation is the n data closest to the current time, and n is an integer preset by the operator. The value must be representative and should cover at least 30 days of data at the same time. The winding loss sequence H2 is processed in the same and synchronous manner as described above to obtain the mean winding loss μp_i. Then, based on the mean group loss μp_i and the mean total active power loss μh_i, hi'=hi-μh_i and pi'=pi-μp_i are executed to perform detrending operations on the total active power loss sequence H1 and the winding loss sequence H2, respectively. Finally, the total active power loss during detrending is obtained as H1'=h1',h2',...,hj' and the detrending winding loss sequence is obtained as H2'=p1',p2',...,pj'. It should be noted that the total active power loss and winding loss of a transformer are affected by the load cycle, exhibiting a clear daily cycle pattern, such as the daily peak and valley of industrial electricity consumption and the influence of day and night temperature differences on resistance. By subtracting the historical average at the same time on the same day, the periodic background trend is extracted. The remaining component mainly reflects abnormal fluctuations that deviate from the normal pattern, avoiding the normal daily cycle changes from masking the fault characteristics, and directly reflecting the deviation of the current loss from the normal baseline.

[0028] Next, obtain the scale parameter of the detrended total active power loss H1', ​​and denote it as the first scale parameter Ch1; The scale parameter of the detrended winding loss sequence H2' is obtained synchronously and denoted as the second scale parameter Ch2; It should be noted that the first scale parameter Ch1 is at least one of the standard deviation, mean absolute deviation, or interquartile range of the detrended total active power loss H1', ​​and the second scale parameter Ch2 is similar. Next, normalization processing is performed on H1' and H2' using gi=hi' / Ch1 and qi=pi' / Ch2 respectively, resulting in two data sequences after normalization, denoted as the normalized total active power loss sequence G1=g1,g2,...,gj and the normalized winding loss sequence G2=q1,q2,...,qj respectively.

[0029] Example 4

[0030] This embodiment, based on embodiment 3, further discloses a point-to-point trajectory mapping method using a sliding window mean decentralization method, such as... Figure 2 As shown, it specifically includes the following: Obtain the sampling time sequence from the normalized total active power loss sequence G1=g1,g2,...,gj and the normalized winding loss sequence G2=q1,q2,...,qj, i.e. time t1 to the j-th sampling time; Then extract the sliding window W and the sliding step size Δ. Both the sliding window W and the sliding step size Δ are preset positive integers. The sliding window W is taken as a reference value of 60 sampling points, corresponding to one hour. The sliding step size Δ is half of the sliding window W, that is, it slides once every 30 minutes. Next, starting from the detection start time t1, the sliding window is determined sequentially according to the sliding step size Δ. The total number of determined sliding windows is then counted, denoted as K, where K is not a constant value and increases over time. The last sliding window waits for the time to accumulate 60 sampling points. This indicates rounding down; +1 indicates the last sliding window. For any kth sliding window among the K sliding windows determined in the above operation, determine the sampling time covered by the kth sliding window, and determine the sampling time range as [t_min, t_max], where t_min is the start sampling time of the kth sliding window and t_max is the end sampling time of the kth sliding window; Next, the normalized total active power loss corresponding to all sampling times within the sampling time range [t_min, t_max] is determined from the normalized total active power loss sequence G1, and mean processing is performed to obtain a mean value, which is marked as the sliding window mean μG1_k associated with the normalized total active power loss sequence G1 within the sampling time range [t_min, t_max] (i.e. the k-th sliding window), k=1,2,...,K; Similarly, based on the above steps, the normalized winding loss sequence G2 is synchronized to determine the sliding window mean μG2_k; Obtain the normalized total active power loss gi and normalized winding loss qi corresponding to any sampling time i within the sampling time range [t_min, t_max] from the normalized total active power loss sequence G1 and the normalized winding loss sequence G2. The abscissa Xi is constructed using Xi=gi-μG1_k, the ordinate Yi is constructed using Yi=qi-μG2_k, and a pair of points (Xi,Yi) is generated based on the constructed abscissa Xi and ordinate Yi. Based on the above method, synchronization processing is performed on all sampling times [t_min, t_max] within the k-th sliding window to obtain all point pairs, which are arranged in the order of sampling times and denoted as the point pair set DH_k; By applying the same processing method to K sliding windows, we finally obtain a set of K point pairs DH_1,DH_2,...,DH_K.

[0031] Based on the above steps, extract the set of point pairs DH_k associated with any k-th sliding window; A two-dimensional coordinate system is constructed with the local deviation of total active power loss as the horizontal axis and the local deviation of winding loss as the vertical axis (the coordinate axes represent the relative quantities after removing the mean, i.e., the local deviations). Any pair of points (Xi, Yi) in the point pair set DH_k is aligned with the horizontal and vertical axes of the two-dimensional coordinate system, and the corresponding data points are plotted. Finally, all data points are plotted in the two-dimensional coordinate system based on the point pair set DH_k, and all data points are connected in the order of sampling time to obtain the discrete trajectory curve, denoted as LS_k. It is important to note that since Δ is less than W, the same sampling moment will appear in multiple windows. Each window will calculate different point pairs. Because the mean values ​​are different, the point pairs will appear repeatedly in multiple trajectories with different coordinates. To address this issue, in the actual implementation of this scheme, a local trajectory is generated independently for each window to analyze the dynamic behavior within that time period. Therefore, each local trajectory corresponds to an independent local time segment, reflecting the relative relationships within that segment, without needing to consider the issue of different point pairs.

[0032] Example 5

[0033] This embodiment, based on embodiment 4, further discloses a method for detecting transformer loss loop anomalies based on the sector interval polarity boundary and arrangement entropy, such as... Figure 3 As shown, it specifically includes the following: First, taking the origin of the two-dimensional coordinate system constructed in Example 4 as the pole, the plane containing the two-dimensional coordinate system is divided into M sector intervals according to the preset angle step Δθ∈[1°,5°], i.e. M=2π / Δθ; For any discrete trajectory curve LS_k associated with the kth sliding window, traverse all data points on it, filter out the data points whose polar corners are located in the corresponding sector interval, calculate the polar radius of each data point, and determine the maximum polar radius R_max and the minimum polar radius R_min in the corresponding sector interval respectively. It should be noted that if there are no data points in a certain sector interval, linear interpolation is performed based on the maximum extreme diameter R_max and minimum extreme diameter R_min of the adjacent sector intervals of the corresponding sector interval, respectively, as the estimated maximum and minimum extreme diameters of the sector interval. The maximum number of consecutive interpolation sector intervals shall not exceed 3. If more than 3 are exceeded, the data is marked as invalid and the sector interval is skipped.

[0034] Next, calculate the difference between the maximum and minimum extreme radii within all sector intervals corresponding to the k-th sliding window, and mark it as the radial width. Arrange them in angular order to form the radial width sequence Z1, Z2, ..., ZM. The angular order starts with the positive direction of the vertical axis and proceeds clockwise. Radial widths Z1 and ZM are physically adjacent, and subsequent entropy calculations will perform splicing processing on the beginning and end of the sequence.

[0035] Then, the radial width sequence Z1, Z2, ..., ZM is reconstructed in phase space. Based on the preset embedding dimension d and time delay τ, the radial width sequence is transformed into several radial width subsequences, where the embedding dimension d satisfies M-(d-1)τ≥100.

[0036] The permutation patterns of numerical values ​​in each radial width subsequence are statistically analyzed, and the probability of each permutation pattern is calculated to obtain the permutation entropy ET_k of the k-th sliding window, which is ET_k = -∑(Q×log(Q)), where Q is the probability of any permutation pattern. Finally, the permutation entropy ET_k is used as the loss ring anomaly index of the k-th sliding window. Obtain historical normal data of several electrical steel strip transformers of the same specification and model as the electrical steel strip transformers described in this solution under normal operating conditions, and lock the baseline range of the permutation entropy based on these historical normal data. Calculate the permutation entropy distribution by taking the 2.5%-97.5% quantile to obtain the lower limit threshold ET_min and the upper limit threshold ET_max. If the current sliding window's loss loop anomaly index ET_k < ET_min or ET_k > ET_max, then the current sliding window is determined to be abnormal, and an anomaly alarm signal is output.

[0037] The permutation entropy of the sliding window is continuously determined and trend analysis is performed. If the permutation entropy increases monotonically for more than the preset duration of r sliding windows, and the cumulative increase of the permutation entropy within r sliding windows exceeds the preset percentage α%, a performance degradation warning is generated. The reference value of α% is 20%, and operators can adjust it according to actual needs. When an anomaly or performance degradation warning is triggered, an anomaly alarm signal is output.

[0038] Example 6

[0039] This embodiment is implemented based on Embodiment 5. The difference between Embodiment 5 and Embodiment 5 is that this embodiment does not use fixed sector interval division parameters and fixed permutation entropy thresholds. Instead, it introduces an adaptive parameter adjustment mechanism based on the running state. Specifically, this includes the adaptive determination of the number of sector intervals M and the phase space reconstruction parameters: embedding dimension d, and time delay τ, as follows: First, determine the total number of data point pairs within the k-th sliding window, denoted as N_k; Using M_k=round(β×N_k) (1 / 3) The number of sector intervals is determined, where β is an empirical coefficient with a value range of [2,5]. In this embodiment, β=3 is preferred. This method is based on a variant of Sturges' rule in statistics, ensuring that there are enough data points on average in each sector interval for estimating the polar radius boundary, avoiding too many data-free intervals due to overly dense sector intervals or too sparse intervals leading to a decrease in accuracy. Next, for the radial width sequence Z1, Z2, ..., ZM_k of the k-th sliding window, the optimal time delay τ_k and embedding dimension d_k are determined using the mutual information method. The time delay τ_k is calculated by using the mutual information function of the radial width sequence Z1, Z2, ..., ZM_k, and selecting the point where the mutual information function first decreases to its maximum value. The delay corresponding to the time multiplication is τ_k, and ℮ is the natural logarithm; The embedding dimension d_k adopts the pseudo nearest neighbor method FNN. When the FNN ratio first falls below a preset threshold of 1%, the corresponding dimension is used as the embedding dimension d_k.

[0040] Next, the number of sector intervals M_k is locked based on the total number of data point pairs N_k; With the origin of the coordinate system as the pole, the plane is divided into M_k sector intervals. All data points are traversed to determine the maximum radius R_max and minimum radius R_min of each interval. For intervals with no data points, the linear interpolation strategy described in Example 5 is executed. If the number of consecutive invalid intervals exceeds 10% of M_k, then M_k is reduced to M_k-2 and this step is repeated. Then, the time delay τ_k and the embedding dimension d_k are extracted, and the radial width sequence is reconstructed in phase space to calculate the permutation entropy ET_k.

[0041] When the total number of data point pairs N_k is less than the preset minimum sample size threshold, such as N_k < 50, the system will automatically revert to the fixed parameter mode in Example 5 and issue a corresponding automatic revert prompt message.

[0042] By adopting adaptive parameters, the robustness of this solution under different load fluctuations is improved. In the simulation test, for the scenario where the load rate fluctuates drastically from 20% to 120%, when using the fixed parameters in Example 5, about 15% of the sliding windows have a deviation of more than ±10% in the permutation entropy calculation due to uneven distribution of data points. However, after adopting the adaptive parameters of this example, the deviation is controlled within ±5%. Meanwhile, for slowly evolving inter-turn short-circuit faults, the adaptive phase space parameters can more sensitively capture changes in the dynamic structure of the radial width sequence, ultimately improving the anomaly detection time by an average of 2 to 3 sliding window cycles compared to the fixed parameter scheme.

[0043] All data in the formulas described above have been calculated with dimensions removed. Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0044] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0045] It should be stated that all user data collected in this application was collected with the user's consent and authorization. Furthermore, the uses of user data are legal and compliant, and the use and processing of user data comply with the relevant laws, regulations, and standards of the relevant regions.

Claims

1. An adaptive detection method for electrical steel strip transformers, characterized in that, The method includes: Step 1: Lock the input power and output power in a time-aligned manner to generate the total active power loss sequence H1 associated with the time sequence, and simultaneously lock the load current and winding resistance in real time to generate the winding loss sequence H2 corresponding to the total active power loss sequence. Step 2: Based on the historical average values ​​of H1 and H2 at the corresponding time, detrending is performed on each of them. Normalization is then performed using the scale parameters of the two sequences after detrending to generate the normalized total active power loss sequence G1 and the normalized winding loss sequence G2. Step 3: Lock the mean values ​​of sliding windows G1 and G2 within the sliding window respectively. Use the different mean values ​​of sliding windows and the data contained in G1 and G2 within different sliding windows to perform point pair extraction, construct a set of point pairs, and plot the discrete trajectory curve in the pre-constructed two-dimensional coordinate system. Step four: Based on the discrete trajectory curve, perform angular division to determine several sector intervals, determine the maximum and minimum extreme radii based on the extreme radii boundaries and angular step size of the sector intervals, further evaluate the loss loop anomaly index of the corresponding sliding window, lock the abnormal state in real time, and generate an abnormal alarm signal.

2. The method according to claim 1, characterized in that, In step one, the specific method for generating the time-series-related total active power loss sequence H1 is as follows: Taking the current time as the start time of detection, denoted as t1, the input power and output power of the electrical steel strip transformer are collected at a preset sampling frequency to obtain the input power sequence P_in1, P_in2, ..., P_inj and the output power sequence P_out1, P_out2, ..., P_outj, where j is the total number of samples, which increases with time. Calculate the difference between input power and output power at each sampling time to determine the total active power loss sequence H1=h1,h2,...,hj.

3. The method according to claim 1, characterized in that, In step one, the specific method for generating the winding loss sequence H2 corresponding to the total active power loss sequence is as follows: Determine the load current sequence D1, D2, ..., Dj and winding resistance R corresponding to the input power sequence and output power sequence timing. The winding resistance R is taken as the average resistance value of the electrical steel strip transformer under normal operating conditions and is considered as a known value. Using pi = v × (Di) 2 ×R calculates the winding loss at each sampling time, forming a winding loss sequence H2=p1,p2,...,pj, where v is the total number of phase windings of the electrical steel strip transformer, i is the counting index, and 1≤i≤j.

4. The method according to claim 1, characterized in that, In step two, the specific methods for performing detrending based on the historical averages of H1 and H2 at corresponding times are as follows: The total active power loss sequence H1 = h1, h2, ..., hj is analyzed. Similarly, the winding loss sequence H2 = p1, p2, ..., pj is obtained; For any sampling time i, lock its corresponding time within the day, and perform an average of the total active power loss of the electrical steel strip transformer at the same time within the day under the historical normal operation state to obtain the average total active power loss μh_i, where the averaged total active power loss is the n closest to the current time, and n is a preset integer. Similarly, determine the mean winding loss μp_i; Detrending is performed on the total active power loss sequence H1 and the winding loss sequence H2 using hi'=hi-μh_i and pi'=pi-μp_i respectively, to obtain the detrended total active power loss H1'=h1',h2',...,hj' and the detrended winding loss sequence H2'=p1',p2',...,pj'.

5. The method according to claim 4, characterized in that, In step two, the specific method for performing normalization using the scale parameters of the two sequences after detrending is as follows: Obtain the scale parameter of the detrended total active power loss H1', ​​denoted as the first scale parameter Ch1, and obtain the scale parameter of the detrended winding loss sequence H2', denoted as the second scale parameter Ch2. The first scale parameter Ch1 is at least one of the standard deviation, mean absolute deviation, or interquartile range of the detrended total active power loss H1', ​​and the second scale parameter Ch2 is similar. Normalization of H1' and H2' is performed using gi=hi' / Ch1 and qi=pi' / Ch2 respectively, resulting in the normalized total active power loss sequence G1=g1,g2,...,gj and the normalized winding loss sequence G2=q1,q2,...,qj.

6. The method according to claim 5, characterized in that, In step three, point pair extraction is performed using different sliding window mean values ​​and the data contained in G1 and G2 within different sliding windows. The specific method for constructing the point pair set is as follows: Extract the sliding window W and the sliding step size Δ, where W and Δ are both preset positive integers, and Δ≤W; Starting from the detection start time t1, the sliding windows are sequentially slid with a sliding step size Δ to obtain K sliding windows, where, Indicates rounding down; For any k-th sliding window, take the sampling time range [t_min, t_max] it covers, and determine the normalized total active power loss corresponding to [t_min, t_max] from the normalized total active power loss sequence G1. Calculate the mean, denoted as the sliding window mean μG1_k, k=1,2,...,K; Similarly, the normalized winding loss sequence G2 is processed to determine the sliding window mean μG2_k; Take the normalized total active power loss gi and normalized winding loss qi corresponding to any sampling time i within the sampling time range [t_min, t_max]. The x-coordinate Xi and y-coordinate Yi are constructed using Xi=gi-μG1_k and Yi=qi-μG2_k respectively, generating a pair of points (Xi,Yi); Similarly, synchronization processing is performed on all sampling times [t_min, t_max] within the k-th sliding window to obtain the point pair set DH_k; Similarly, processing the K sliding windows yields a set of K point pairs: DH_1, DH_2, ..., DH_K.

7. The method according to claim 6, characterized in that, In step three, the specific method for plotting the discrete trajectory curve in the pre-constructed two-dimensional coordinate system is as follows: Extract the set of point pairs DH_k associated with any k-th sliding window; A two-dimensional coordinate system is constructed with the local deviation of total active power loss as the horizontal axis and the local deviation of winding loss as the vertical axis. Get any pair of points (Xi, Yi) from the set of point pairs DH_k, align them with the horizontal and vertical axes of the two-dimensional coordinate system, and plot the data points; Similarly, based on the point pair set DH_k, all data points are plotted in a two-dimensional coordinate system, and all data points are connected in the order of sampling time to obtain the discrete trajectory curve, denoted as LS_k.

8. The method according to claim 1, characterized in that, In step four, the specific method for determining the maximum and minimum polar radii based on the polar radius boundary and angular step size of the sector interval is as follows: With the origin of the two-dimensional coordinate system as the pole, the two-dimensional coordinate system is divided into M sector intervals according to a preset angular step Δθ, where M=2π / Δθ, Δθ∈[1°,5°]; For each sector interval, traverse all data points on the discrete trajectory curve LS_k, filter out the data points where the polar corners are located in the corresponding sector interval, calculate the polar radius of each data point, and record the maximum polar radius R_max and the minimum polar radius R_min in the corresponding sector interval. If there are no data points in a certain sector interval, then the linear interpolation of the maximum extreme diameter R_max and the minimum extreme diameter R_min of the adjacent sector interval is taken as the estimated maximum and minimum extreme diameters of that sector interval.

9. The method according to claim 8, characterized in that, In step four, the specific method for further evaluating the loss loop anomaly index of the corresponding sliding window, locking the abnormal state in real time, and generating an anomaly alarm signal is as follows: For any k-th sliding window, extract the difference between the maximum and minimum extreme radii of each sector interval, denoted as the radial width, and arrange them in angular order to form the radial width sequence Z1, Z2, ..., ZM; The radial width sequence Z1, Z2, ..., ZM is reconstructed in phase space. Based on the preset embedding dimension d and time delay τ, the radial width sequence is transformed into several radial width subsequences. The permutation patterns of numerical values ​​in each radial width subsequence are statistically analyzed, and the probability of each permutation pattern is calculated to obtain the permutation entropy ET_k of the k-th sliding window, which is used as the loss ring anomaly index of the k-th sliding window. Based on historical normal data of electrical steel strip transformers of the same specification and model under normal operating conditions, the baseline range of the permutation entropy is locked, and the permutation entropy distribution is calculated to obtain the lower limit threshold ET_min and the upper limit threshold ET_max. If the current sliding window's loss loop anomaly index ET_k < ET_min or ET_k > ET_max, then the current sliding window is determined to be abnormal, and an anomaly alarm signal is output.

10. The method according to claim 9, characterized in that, Step four also includes simultaneously performing trend analysis on the arrangement entropy of the continuous sliding windows. If the arrangement entropy increases continuously for more than the duration of a preset r sliding windows and the cumulative increase exceeds a preset percentage α, a performance degradation warning is generated. When an anomaly or performance degradation warning is triggered, an anomaly alarm signal is output.

Images