A time-frequency domain evidence fusion method, apparatus, electronic device, and storage medium for detecting harmonic variations in non-stationary harmonic data.

Abstract

This invention discloses a time-frequency domain evidence fusion method, apparatus, electronic device, and storage medium for detecting harmonic variations in non-stationary harmonic data, belonging to the field of electrical digital data processing technology. The method involves performing a Fast Fourier Transform on the original current waveform time-domain data; dividing the original current waveform time-domain data into subsequences and calculating the nonlinear weighted normalized Euclidean distance; calculating the posterior probability distribution of the time intervals between change points and recording the number of times each time point is inferred to be a change point; normalizing the nonlinear weighted matrix contour and the number of times each time point is inferred to be a change point to obtain sequences arranged by time; truncating subsequences and calculating shape similarity to obtain the intervals where change points exist; selecting candidate times exceeding a threshold as mutually exclusive identification frames and constructing a synthetic basic probability allocation function; calculating the trust function and likelihood function, and obtaining the final change point using a discriminant formula. This invention can accurately and effectively find change points.

Description

Technical Field

[0001] This invention relates to the field of electrical digital data processing technology, and in particular to a time-frequency domain evidence fusion method, apparatus, electronic device, and storage medium for detecting harmonic variable operating conditions applicable to non-stationary harmonic data. Background Technology

[0002] With the widespread integration of high-proportion power electronic devices and renewable energy into the power grid, the harmonic problem in the power system is becoming increasingly severe. Harmonic pollution not only affects power quality but also threatens the safe operation of equipment, becoming a significant hidden danger in the power system. Harmonic source modeling, as a prerequisite for assessing harmonic hazards and formulating harmonic suppression strategies, is inseparable from the detection of changing operating conditions (changing points), which is crucial for maintaining qualified power quality and ensuring the safe operation of equipment.

[0003] "Change of operating condition" refers to the moment when a time series undergoes a structural change from one state distribution to another. Unlike data abrupt changes caused by sudden noise or disturbances, structural change points arise from changes in the data distribution of the time series. Change point detection mechanisms are generally based on probability distribution tests of the time series. The process of change point detection requires not only observation data before the time point being measured but also data after the time point. Change point detection methods based on the Cumulative Sum (CUSUM) test are widely used. The CUSUM test assumes that the probability distribution of the time series changes before and after a change, and it constructs likelihood ratios before and after the change. When a signal undergoes a structural change and a positive shift occurs, it continuously increases, which is a cumulative process. When the accumulation reaches a certain level (greater than a set threshold), the time series can be considered to have changed. Alexander et al. described the detection mechanism of the CUSUM test for different changes in time series. The CUSUM test can accumulate time series changes, and a decision is ultimately made through a threshold test. Since fixed thresholds are typically used in CUSUM tests, adaptive threshold decision-making has become an important aspect of CUSUM research. Alippi et al. introduced a global confidence factor based on the basic CUSUM test method, achieving adaptive parameter calculation. Similarly, Verdier et al. introduced new constraints to the final decision, replacing the original average run length constraint, to achieve adaptive decision-making at the point of change. Some literature, combining local data features, proposed a data-driven threshold based on the steepest descent point, replacing the traditional CUSUM statistic with a shape-based statistic, and further identified the point of change through unimodal identification. Furthermore, research on change point detection based on Bayesian tests typically treats relevant parameters in the model, including the point of change, as random variables. First, the prior distribution of the parameters is obtained using prior knowledge. Then, based on the data information of the sample sequence, the posterior distribution of the parameters at the point of change is calculated using Bayes' theorem, obtaining the point with the highest probability among the possible values ​​of the parameters as an estimate of the time of change.

[0004] Besides the aforementioned statistically based change point detection research, numerous change point detection methods have been proposed in other fields. One such method structurally summarizes multivariate time series change point detection algorithms into three elements: cost function, search method, and constraint on the number of change points, with each element described and discussed separately. Other methods first estimate the model's regression function using local linearity, then construct monitoring statistics based on wavelet coefficients, derive the asymptotic distribution of the monitoring statistics under both null and alternative hypotheses, and subsequently present an online bootstrap monitoring method.

[0005] Therefore, in the face of the problems that existing models are not good at capturing the time-varying characteristics of non-stationary harmonic data distribution and have weak generalization ability, there is a need for a time-frequency domain evidence fusion method for harmonic variable condition detection that is suitable for non-stationary harmonic data to construct a real-time variable point detection statistic that can accurately characterize the harmonic variation characteristics and alleviate the update of distribution drift. Summary of the Invention

[0006] The purpose of this invention is to propose a time-frequency domain evidence fusion method, apparatus, electronic device, and storage medium for detecting harmonic variable operating conditions applicable to non-stationary harmonic data.

[0007] A time-frequency domain evidence fusion method for harmonic variable operating condition detection applicable to non-stationary harmonic data includes:

[0008] A fast Fourier transform is performed on the time-domain data of the sampled raw current waveform to obtain frequency-domain data containing harmonics.

[0009] The original current waveform time-domain data is divided into subsequences using a sliding window. The nonlinear weighting method is used to calculate the nonlinear weighted normalized Euclidean distance between each subsequence and all other subsequences at each time step, and the nonlinear weighting matrix profile is obtained.

[0010] Based on frequency domain data containing harmonics, calculate the posterior probability distribution of the time interval between change points to obtain the change point location that can be determined at the change point time. Traverse the frequency domain data and record the number of times each time point is inferred to be a change point.

[0011] An adaptive fuzzy membership function is used to normalize the nonlinear weighted matrix contour and the number of times it is inferred to be a change point at each time step, resulting in a sequence arranged by time step. Subsequences are extracted, and the shape similarity between subsequences is calculated. The shape similarity is traversed, and the threshold method is used to obtain the interval where the change point exists.

[0012] Based on the interval where the change point exists, candidate moments exceeding the threshold are selected as mutually exclusive recognition frameworks. Time-domain basic probability allocation functions and frequency-domain basic probability allocation functions are constructed respectively, and Dempster synthesis rules are used to fuse them to obtain a synthetic basic probability allocation function.

[0013] Calculate the confidence function and likelihood function to obtain the lower and upper bounds of the confidence interval for each candidate time. Based on the lower and upper bounds of the confidence interval for each candidate time, obtain the discrimination formula, and then obtain the final change point from the discrimination formula.

[0014] Furthermore, the profile of the nonlinear weighted matrix is ​​the minimum of the nonlinear weighted normalized Euclidean distance.

[0015] Furthermore, the formula for calculating the number of times each moment is inferred to be a change point is:

[0016] ;

[0017] Where IN[t] represents the number of times a point is predicted to be a change point at time t. , where i is the time label. is the most likely running length at time i, m is the length of the subsequence, n is the time number, and I(.) is the indicator function, which takes the value 1 when the value in the parentheses is true, and 0 otherwise.

[0018] Furthermore, the discriminant formula is:

[0019] ;

[0020] in, For the final change point, For mutually exclusive identification frameworks, and These represent the lower and upper limits of the trust range of the identification framework, respectively.

[0021] A time-frequency domain evidence fusion harmonic variable condition detection device suitable for non-stationary harmonic data includes:

[0022] The frequency domain transformation module is used to perform a fast Fourier transform on the sampled original current waveform time domain data to obtain frequency domain data containing harmonics.

[0023] The distance calculation module is used to divide the original current waveform time-domain data into subsequences using a sliding window, and to calculate the nonlinear weighted normalized Euclidean distance between each subsequence and all other subsequences at each time step using a nonlinear weighting method, thereby obtaining the nonlinear weighted matrix profile.

[0024] The change point recording module is used to calculate the posterior probability distribution of the time interval between change points based on frequency domain data containing harmonics, obtain the change point location that can be determined at the change point time, traverse the frequency domain data, and record the number of times each time point is inferred to be a change point.

[0025] The interval acquisition module is used to normalize the nonlinear weighted matrix contour and the number of times it is inferred to be a change point at each time step using an adaptive fuzzy membership function, thereby obtaining a sequence arranged by time step, extracting subsequences, calculating the shape similarity between subsequences, traversing the shape similarity, and using a threshold method to obtain the interval where the change point exists.

[0026] The function construction module is used to select candidate moments exceeding the threshold as mutually exclusive recognition frames based on the interval where the change point exists. It constructs the time-domain basic probability allocation function and the frequency-domain basic probability allocation function respectively, and uses the Dempster synthesis rule to fuse them to obtain the synthetic basic probability allocation function.

[0027] The change point discrimination module is used to calculate the confidence function and the likelihood function, obtain the lower limit and upper limit of the confidence interval for each candidate time, obtain the discrimination formula based on the lower limit and upper limit of the confidence interval for each candidate time, and obtain the final change point from the discrimination formula.

[0028] Furthermore, the nonlinear weighted matrix profile in the distance calculation module is the minimum value of the nonlinear weighted normalized Euclidean distance.

[0029] Furthermore, the formula for calculating the number of times a point is inferred to be a change point at each moment in the change point recording module is as follows:

[0030] ;

[0031] Where IN[t] represents the number of times a point is predicted to be a change point at time t. , where i is the time label. is the most likely running length at time i, m is the length of the subsequence, n is the time number, and I(.) is the indicator function, which takes the value 1 when the value in the parentheses is true, and 0 otherwise.

[0032] Furthermore, the discrimination formula in the change point discrimination module is:

[0033] ;

[0034] in, For the final change point, For mutually exclusive identification frameworks, and These represent the lower and upper limits of the trust range of the identification framework, respectively.

[0035] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps of a time-frequency domain evidence fusion method for detecting harmonic variable operating conditions applicable to non-stationary harmonic data.

[0036] A storage medium storing a computer program that, when executed by a processor, implements the various steps of a time-frequency domain evidence fusion method for detecting harmonic variable operating conditions applicable to non-stationary harmonic data.

[0037] The beneficial effects of this invention are as follows:

[0038] 1. The method of this invention can fully mine the operating condition change information in the time and frequency domains of harmonic data, and use evidence fusion theory to infer the abrupt change time of the harmonic data distribution, thereby reducing the false alarm rate. The constructed operating condition change point discrimination formula can accurately identify all change points.

[0039] 2. In the process of mining harmonic time-domain information, this invention uses nonlinear weighted normalized Euclidean distance to calculate the nonlinear weighted matrix profile, which solves the problem that traditional matrix profiles cannot identify repeated operating condition changes. It can accurately capture waveform changes and thus find operating condition change points, improving model robustness. In the process of mining harmonic frequency-domain information, an online detection algorithm is used to continuously update the running length probability, and the number of times is used instead of the running length as the basis for judging change points. This solves the problem that it cannot identify when the running length is close to zero, effectively improving detection accuracy and avoiding omissions.

[0040] 3. This invention can divide non-stationary harmonic data sequences into multiple quasi-steady-state periods by accurately determining change points. Based on the abrupt changes (change points) in the distribution of non-stationary harmonic data detected by the time-frequency domain evidence fusion algorithm, the continuous interval between change points constitutes a quasi-steady-state period, laying the foundation for subsequent high-precision segmented modeling. Segmented modeling can avoid the decrease in model generalization ability caused by direct modeling of non-stationary data and eliminate the interference of heterogeneous data. Attached Figure Description

[0041] Figure 1 This is a flowchart of the time-frequency domain evidence fusion method for detecting harmonic variations in non-stationary harmonic data, applicable to the present invention.

[0042] Figure 2 (a) is the effective value of the voltage; Figure 2 (b) is the effective value of the current; Figure 2 (c) represents the effective value of the harmonic voltage; Figure 2 (d) represents the effective value of the harmonic current.

[0043] Figure 3 (a) is the original effective value of the current; Figure 3 (b) represents the MP value of the matrix contour; Figure 3 (c) is the NWMP value; Figure 3 (d) represents the probability distribution of the running length; Figure 3 (e) represents the most likely run length; Figure 3 (f) represents the number of times IN is inferred to be a change point; Figure 3 (g) represents the shape similarity between the NWMP and the number of inferred dislocation points.

[0044] Figure 4 This is a schematic diagram of the time-frequency domain evidence fusion harmonic variable operating condition detection device applicable to non-stationary harmonic data according to the present invention.

[0045] Figure 5 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0046] This invention proposes a time-frequency domain evidence fusion method, apparatus, electronic device, and storage medium for detecting harmonic variable operating conditions applicable to non-stationary harmonic data. The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0047] Figure 1 This is a flowchart of the time-frequency domain evidence fusion method for harmonic variable operating condition detection applicable to non-stationary harmonic data according to the present invention; specifically:

[0048] A fast Fourier transform is performed on the time-domain data of the sampled raw current waveform to obtain frequency-domain data containing harmonics. Figure 2 (a) is the effective value of the voltage; Figure 2 (b) is the effective value of the current; Figure 2 (c) represents the effective value of harmonic voltage, where U3 is the 3rd harmonic voltage, U5 is the 5th harmonic voltage, U7 is the 7th harmonic voltage, U9 is the 9th harmonic voltage, U11 is the 11th harmonic voltage, U13 is the 13th harmonic voltage, U15 is the 15th harmonic voltage, and U17 is the 17th harmonic voltage. Figure 2 (d) represents the effective value of the harmonic current. I3 is the 3rd harmonic current, I5 is the 5th harmonic current, I7 is the 7th harmonic current, I9 is ​​the 9th harmonic current, I11 is the 11th harmonic current, I13 is the 13th harmonic current, I15 is the 15th harmonic current, and I17 is the 17th harmonic current.

[0049] Given an original current waveform sequence I = {i(1), i(2), ..., i(n)}, representing the current data from time 1 to time n, a parameter m is set, which is the ratio of the sampling rate to 50 Hz. The sampling rate is the number of sampling points per unit time (unit: Hz). In general applications, the sampling rate ranges from 1 to 10 kHz. A sliding window is used to divide the original sequence into n-m+1 subsequences of length m. Here, I... t;m ={i(t),i(t+1),…,i(t+m-1)} is defined as the subsequence corresponding to time t. The formula for calculating the nonlinear weighted normalized Euclidean distance between the subsequence at time t and any other subsequence at time j is:

[0050] ;

[0051] in, This represents the nonlinear weighted standard Euclidean distance between the subsequences at time t and at time j. , and m is the length of the subsequence, Q=I t;m , Indicate Q and I j;m The inner product, I j;mLet be the subsequence corresponding to time j. This represents the mean of Q. This represents the standard deviation of Q. Indicate I j;m The mean, Indicate I j;m The standard deviation. Indicate Q and I j;m Non-linear weights between them The calculation formula is:

[0052] ;

[0053] in, This represents the time interval between two subsequences. The baseline value representing the non-linear weight. The scale parameter represents the time offset. The scale parameter controls the amplitude of the distribution function, that is, the change in the "steepness" of the curve as it grows or decreases.

[0054] Define the profile of the nonlinear weighted matrix at time t as NWMP[t], and its calculation formula is as follows:

[0055] NWMP[t]=min(d tj );

[0056] Where t is fixed, I j;m Take all subsequences in I except for time t, and then calculate d for each subsequence step by step. tj min( ) indicates taking the minimum value among them.

[0057] For multidimensional harmonic data obtained after Fast Fourier Transform Where n-m+1 is the number of frequency domain samples, Each sample is represented by d harmonic quantities of interest. The latent variable r is defined. t This represents the possible running length at time t, that is, the time interval between time t and the previous change point. Generally, let t = 1 be the first point of change, that is, the running length at t = 1. The probability distribution is According to this definition, at time t, r t The possible set of values ​​is redefining This represents the harmonic data at the previous t time steps. The core of this algorithm is to traverse the harmonic data stream and use Bayesian inference to update the conditional distribution. Calculate the conditional distribution at time t. The calculation formula is:

[0058] ;

[0059] in, express and The joint probability distribution of is calculated using the following formula:

[0060] ;

[0061] The conditional probabilities are all obtained recursively. By setting a danger function The formula for its calculation is as follows:

[0062] ;

[0063] ;

[0064] Where P is the probability distribution of the running length, which is a discrete exponential distribution. It indicates a certain time interval.

[0065] Define the most likely running length at time t as The calculation formula is defined as follows:

[0066] ;

[0067] Here, the function argmax(f(x)) represents the variable point x (or the set of x) that makes f(x) reach its maximum value, and j represents the set L. t The elements in.

[0068] Finally, using Infer the position of the previous turning point at time t. After traversing the data once, record the number of times each time point is inferred to be a change point. Define IN[t] as the number of times time t is inferred to be a change point during the traversal, and its calculation formula is:

[0069] ;

[0070] Where i is the time label. , This is an indicator function; it takes the value 1 when the value inside the parentheses is true, and 0 otherwise. IN will have a larger value at points of change.

[0071] Since the nonlinear weighted matrix profile NWMP[t] at time t and the frequency domain variable point detection index IN[t] have different dimensions, an adaptive fuzzy membership function is used. The sequences NWMP={NWMP[1], NWMP[2],…, NWMP[n-m+1]} and IN={IN[1], IN[2],…, IN[n-m+1]} are transformed to the range [0,1]. The formula for calculating the adaptive fuzzy membership function is as follows:

[0072] ;

[0073] in, The scale parameter is represented by u, which represents the time offset. The value of the original change point detection index (time domain index or frequency domain index) at a certain time t will be normalized to the interval [0, 1] through the fuzzy membership function.

[0074] The normalized time-domain index sequence NWMP and the frequency-domain index sequence IN are denoted as sequences respectively. and Set the window width w, in Extract the corresponding subsequence at each time step and The subsequence length is w. The shape similarity between two subsequences at the same time t is calculated. The size of the shape similarity is used to identify the interval where the change point is located. The formula for calculating the shape similarity at time t is as follows:

[0075] ;

[0076] in, k is any integer between -w and w, the function The calculation formula is:

[0077] ;

[0078] ;

[0079] It calculates the Pearson correlation coefficient between the two subsequences inside the brackets, and then calculates the shape similarity between the two subsequences at the same time t.

[0080] Let w represent the (l+k)th element in a subsequence of length w. Let represent the lk-th element in a subsequence of length w. Let l represent the l-th element in a subsequence of length w. , Both represent the position of the last element that can be taken from one of the subsequences during the calculation of shape similarity.

[0081] Then, using a sliding window, subsequences at all time points are taken on A and B, and the shape similarity sequence is calculated. Because subsequences with points of change will have a high degree of shape similarity. Therefore, a threshold (such as 3 times the standard deviation) can be set according to specific circumstances. The interval formed by the times when the shape similarity exceeds this threshold is the interval where the change point exists.

[0082] Based on the obtained intervals of change points, all candidate moments exceeding a certain threshold are selected as mutually exclusive recognition frameworks. Each element in this set is called a proposition, K represents the number of candidate variable points, and it is assumed that each identified variable point interval contains only one variable point. The basic probability assignment function is defined on the power set of the identification frame. The probability function on the frame power set It is a mutually exclusive identification framework The set of all subsets, representing the degree of confidence in each proposition, is calculated using the following formula:

[0083] ;

[0084] The sum of the fundamental probability assignment functions for all propositions is 1. Let m1 and m2 represent the fundamental probability assignment functions in the time domain and frequency domain, respectively. The Dempster synthesis rules are as follows:

[0085] ;

[0086] Where α represents the hypothesis that "the change point occurs in a specific set of times"; Represents a subset of propositions supported by non-temporal evidence (NWMP); Let m(α) represent the proposition (subset) supported by frequency domain evidence (Bayesian shift point); m(α) represent the basic probability assignment (BPA) value of the fused evidence for proposition α; m1(β) represent the basic probability assignment (BPA) value of the time domain evidence for proposition β; m2(γ) represent the basic probability assignment (BPA) value of the frequency domain evidence for proposition γ; m represents the basic probability assignment function of the synthesis of m1 and m2; and k' represents the conflict factor, which is calculated as follows:

[0087] ;

[0088] ;

[0089] ;

[0090] in, It is the i-th element in the power set of the identification frame. Let B be the j-th element in the power set of the identification framework, B be the set of all elements in the time domain that belong to the power set, and C be the set of all elements in the frequency domain that belong to the power set.

[0091] The trust function represents the sum of trust levels for all subsets of a proposition. It reflects the degree of direct support for the proposition, and its formula is as follows:

[0092] ;

[0093] The likelihood function represents the degree to which a proposition is not rejected, that is, the sum of the confidence in all possible outcomes of the proposition. The likelihood function provides indirect support for the proposition, and its calculation formula is as follows:

[0094] ;

[0095] According to the DS evidence theory and Representing subsets The lower and upper bounds of a probability distribution are defined by... Representing a subset The uncertain interval. In order to... Find the final change point The constructed discriminant formula is as follows:

[0096] ;

[0097] Finally, for each identified change point interval, the discriminant formula is substituted to determine all change points using this method.

[0098] A current sequence with 2500 sampling points was obtained through numerical simulation, such as... Figure 3 As shown in (a), the actual operating conditions vary at points [0, 500, 1000, 1500, 2000]. Assuming that the fundamental frequency and the 3rd, 5th, and 7th harmonics all follow a normal distribution, the mean and covariance matrix of the current data for each segment are calculated using 500 sampling points as a segment, in order to verify the effectiveness of the present invention. Figure 3 (b) represents the MP value of the matrix contour. Figure 3 (c) represents the NWMP value. Figure 3 (d) represents the probability distribution of the running length. Figure 3 (e) represents the most likely run length. Figure 3 (f) represents the number of times IN is inferred to be a change point. Figure 3 (g) represents the shape similarity between the NWMP and the number of inferred dislocation points.

[0099] The following evaluation indicators are based on four important variable operating condition testing methods, and the actual set of change points is as follows: The estimated set of change points is The lengths of these two sets are respectively and They are not necessarily equal.

[0100] 1. Annotate error indicators

[0101] Annotation error indicator, denoted as , indicating the number of change points in the change point estimation results ( ) and the number of actual change points ( The difference between them is defined as:

[0102] ;

[0103] 2. Hausdorf Distance

[0104] Hausdorff distance (or Hausdorff metric) is denoted as... , representing the maximum matching error between the actual set of change points and the estimated set of change points. Formally, it equals the maximum time distance between the estimated change points and the actual change points, and its calculation formula is:

[0105] ;

[0106] Among them, the first item Indicates the estimated point of change The maximum distance to its nearest real change point. (Second term) Represents the actual point of change The maximum distance to the nearest estimated change point. Finally, the maximum of the two values ​​is taken, representing the most severe local matching error.

[0107] 3. RAND Corporation

[0108] The Rand index is denoted as Rand index. This represents the accuracy of change point detection. The idea is to calculate the accuracy of change point detection by comparing the consistency between two segmentations. First, a set of grouping indices is defined. Non-grouped index set :

[0109] Furthermore, s and t satisfy certain conditions for the segmentation. The determined series of sub-intervals s and t belong to the same subinterval.

[0110] Furthermore, s and t satisfy certain conditions for the segmentation. The determined series of sub-intervals s and t do not belong to the same subinterval.

[0111] Calculate the set of actual change points respectively corresponding and and the estimated set of change points corresponding and By comparing the consistency of the two sets, the Rand index for the change point detection is calculated. The formula is as follows:

[0112] ;

[0113] The RAND index assesses the overall accuracy of change point detection. A RAND index equals 1 when the estimated change point is exactly the same as the actual change point.

[0114] 4. F1 score

[0115] F1 score, denoted as This represents a combined metric of precision and recall in the change point detection results, used to assess the accuracy of the detection algorithm. First, the "set of correct results" is given. The concept, for a real point of change We believe that its successful detection is contingent upon the existence of a predicted change point. Furthermore, if the distance between the two is less than a certain allowable value M, then... The definition of

[0116] ;

[0117] Where M is called the tolerance value, representing the tolerance for prediction deviation. The larger the value of M, the greater the ability to accept prediction errors. According to The definition of precision is as follows:

[0118] ;

[0119] The above formula shows that precision actually means "the proportion of true changes among estimated changes," reflecting the false positive rate of the detection results. The definition of recall is as follows:

[0120] ;

[0121] The above formula shows that recall rate actually means "the proportion of true change points that are successfully detected," reflecting the false negative rate. The F1 score is defined as the harmonic mean of precision and false negative rate, and its calculation formula is:

[0122] ;

[0123] As can be seen from the above formula, the maximum value of the F1 score is 1. At this time, all the real change points have been successfully detected, and there are no results that are incorrectly identified as change points among the estimated change points.

[0124] The following are the comparison results of the evaluation indicators of the evidence fusion of this invention with other detection methods:

[0125]

[0126] Figure 4 This is a schematic diagram of the time-frequency domain evidence fusion harmonic variable operating condition detection device for non-stationary harmonic data according to the present invention. It includes:

[0127] The frequency domain transformation module is used to perform a fast Fourier transform on the sampled original current waveform time domain data to obtain frequency domain data containing harmonics.

[0128] The distance calculation module is used to divide the original current waveform time-domain data into subsequences using a sliding window, and to calculate the nonlinear weighted normalized Euclidean distance between each subsequence and all other subsequences at each time step using a nonlinear weighting method, thereby obtaining the nonlinear weighted matrix profile.

[0129] The change point recording module is used to calculate the posterior probability distribution of the time interval between change points based on frequency domain data containing harmonics, obtain the change point location that can be determined at the change point time, traverse the frequency domain data, and record the number of times each time point is inferred to be a change point.

[0130] The interval acquisition module is used to normalize the nonlinear weighted matrix contour and the number of times it is inferred to be a change point at each time step using an adaptive fuzzy membership function, thereby obtaining a sequence arranged by time step, extracting subsequences, calculating the shape similarity between subsequences, traversing the shape similarity, and using a threshold method to obtain the interval where the change point exists.

[0131] The function construction module is used to select candidate moments exceeding the threshold as mutually exclusive recognition frames based on the interval where the change point exists. It constructs the time-domain basic probability allocation function and the frequency-domain basic probability allocation function respectively, and uses the Dempster synthesis rule to fuse them to obtain the synthetic basic probability allocation function.

[0132] The change point discrimination module is used to calculate the confidence function and the likelihood function, obtain the lower limit and upper limit of the confidence interval for each candidate time, obtain the discrimination formula based on the lower limit and upper limit of the confidence interval for each candidate time, and obtain the final change point from the discrimination formula.

[0133] This embodiment also includes an electronic device and a storage medium. For example... Figure 5As shown, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements various steps in a time-frequency domain evidence fusion method for detecting harmonic variations applicable to non-stationary harmonic data. A storage medium stores a computer program thereon, which, when executed by a processor, implements various steps in a time-frequency domain evidence fusion method for detecting harmonic variations applicable to non-stationary harmonic data.

[0134] As demonstrated by the above embodiments, the method of the present invention can fully mine the operating condition change information in the time and frequency domains of harmonic data, and use evidence fusion theory to infer the abrupt change time of the harmonic data distribution, reducing the false alarm rate and accurately identifying all change points. Furthermore, it solves the problem that traditional matrix contours cannot identify repeated operating condition changes, accurately capturing waveform changes to discover operating condition change points and improving model robustness. In the process of mining harmonic frequency domain information, a Bayesian online detection algorithm is used to continuously update the run length probability, and the number of occurrences is used instead of the run length as the basis for judging change points, solving the problem of inability to identify when the run length is close to zero, effectively improving detection accuracy and avoiding omissions. It also lays the foundation for subsequent high-precision segmented modeling, avoiding the decrease in model generalization ability caused by direct modeling of non-stationary data, and eliminating interference from heterogeneous data.

[0135] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0136] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0137] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0138] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0139] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0140] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A time-frequency domain evidence fusion method for detecting harmonic variations in non-stationary harmonic data, characterized in that, include: A fast Fourier transform is performed on the time-domain data of the sampled raw current waveform to obtain frequency-domain data containing harmonics. The original current waveform time-domain data is divided into subsequences using a sliding window. The nonlinear weighting method is used to calculate the nonlinear weighted normalized Euclidean distance between each subsequence and all other subsequences at each time step, and the nonlinear weighting matrix profile is obtained. Based on frequency domain data containing harmonics, calculate the posterior probability distribution of the time interval between change points to obtain the change point location that can be determined at the change point time. Traverse the frequency domain data and record the number of times each time point is inferred to be a change point. The formula for calculating the number of times each time point is inferred to be a change point is: ； Where IN[t] represents the number of times a point is predicted to be a change point at time t. , where i is the time label. Let m be the most likely running length at time i, m be the length of the subsequence, n be the time number, and I(.) be the indicator function, which takes the value 1 when the value in parentheses is true and 0 otherwise. An adaptive fuzzy membership function is used to normalize the nonlinear weighted matrix contour and the number of times it is inferred to be a change point at each time step, resulting in a sequence arranged by time step. Subsequences are extracted, and the shape similarity between subsequences is calculated. The shape similarity is traversed, and the threshold method is used to obtain the interval where the change point exists. Based on the interval where the change point exists, candidate moments exceeding the threshold are selected as mutually exclusive recognition frameworks. Time-domain basic probability allocation functions and frequency-domain basic probability allocation functions are constructed respectively, and Dempster synthesis rules are used to fuse them to obtain a synthetic basic probability allocation function. Calculate the trust function and likelihood function to obtain the lower and upper limits of the trust interval for each candidate time. Based on the lower and upper limits of the trust interval for each candidate time, obtain the discrimination formula. The final change point is obtained from the discrimination formula. The discriminant formula is: ； in, For the final change point, For mutually exclusive identification frameworks, and These represent the lower and upper limits of the trust range of the identification framework, respectively.

2. The time-frequency domain evidence fusion method for harmonic variable operating condition detection applicable to non-stationary harmonic data as described in claim 1, characterized in that, The nonlinear weighted matrix profile is the minimum value of the nonlinear weighted normalized Euclidean distance.

3. A time-frequency domain evidence fusion harmonic variable operating condition detection device suitable for non-stationary harmonic data, characterized in that, include: The frequency domain transformation module is used to perform a fast Fourier transform on the sampled original current waveform time domain data to obtain frequency domain data containing harmonics. The distance calculation module is used to divide the original current waveform time-domain data into subsequences using a sliding window, and to calculate the nonlinear weighted normalized Euclidean distance between each subsequence and all other subsequences at each time step using a nonlinear weighting method, thereby obtaining the nonlinear weighted matrix profile. The change point recording module is used to calculate the posterior probability distribution of the time interval between change points based on frequency domain data containing harmonics, obtain the change point location that can be determined at the change point time, traverse the frequency domain data, and record the number of times each time point is inferred to be a change point. The formula for calculating the number of times a point is inferred to be a change point at each moment in the change point recording module is as follows: ； Where IN[t] represents the number of times a point is predicted to be a change point at time t. , where i is the time label. Let m be the most likely running length at time i, m be the length of the subsequence, n be the time number, and I(.) be the indicator function, which takes the value 1 when the value in parentheses is true and 0 otherwise. The interval acquisition module is used to normalize the nonlinear weighted matrix contour and the number of times it is inferred to be a change point at each time step using an adaptive fuzzy membership function, thereby obtaining a sequence arranged by time step, extracting subsequences, calculating the shape similarity between subsequences, traversing the shape similarity, and using a threshold method to obtain the interval where the change point exists. The function construction module is used to select candidate moments exceeding the threshold as mutually exclusive recognition frames based on the interval where the change point exists. It constructs the time-domain basic probability allocation function and the frequency-domain basic probability allocation function respectively, and uses the Dempster synthesis rule to fuse them to obtain the synthetic basic probability allocation function. The change point discrimination module is used to calculate the confidence function and the likelihood function, obtain the lower limit and upper limit of the confidence interval for each candidate time, obtain the discrimination formula based on the lower limit and upper limit of the confidence interval for each candidate time, and obtain the final change point from the discrimination formula; The discrimination formula in the change point discrimination module is: ； in, For the final change point, For mutually exclusive identification frameworks, and These represent the lower and upper limits of the trust range of the identification framework, respectively.

4. The time-frequency domain evidence fusion harmonic variable operating condition detection device for non-stationary harmonic data according to claim 3, characterized in that, The nonlinear weighted matrix profile in the distance calculation module is the minimum value of the nonlinear weighted normalized Euclidean distance.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements each step of the time-frequency domain evidence fusion harmonic variable condition detection method for non-stationary harmonic data as described in any one of claims 1 to 2.

6. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step of the time-frequency domain evidence fusion harmonic variable operating condition detection method for non-stationary harmonic data as described in any one of claims 1 to 2.

Images