Structural health monitoring time series signal spike anomaly identification and positioning method and system

Abstract

The present application relates to the technical field of structural health monitoring and engineering signal processing, and provides a structural health monitoring time series signal spike anomaly identification and positioning method and system. The method comprises the following steps: based on adjacent sampling point difference and local robust statistics, candidate edges are extracted and candidate spike events are formed; candidate widths are obtained by fusing multi-source width information after reliability check, and local analysis windows and wavelet scale sets are determined accordingly; multi-scale wavelet analysis is performed in the local analysis window corresponding to the candidate spike event, noise level is estimated scale by scale, and an adaptive threshold is determined; a two-dimensional cross-scale response region is formed according to the adjacency relationship of the super-threshold response in the time and scale directions, the peak time is determined according to the time aggregation response in the response region, and the positioning interval is determined by fusing the time projection of the response region and the candidate edge interval. The method can balance the abnormal spike identification accuracy and the calculation efficiency of multi-channel structural health monitoring data.

Description

Technical Field

[0001] This invention relates to the fields of structural health monitoring and engineering signal processing technology, and in particular to a method and system for identifying and locating anomalies in time-series signals during structural health monitoring. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Structural health monitoring systems continuously collect bridge operational status data over long periods using various sensors, including those for strain, displacement, acceleration, tilt, temperature, and wind speed. During actual operation, factors such as sensor condition, data acquisition equipment performance, power supply fluctuations, communication link quality, electromagnetic interference, environmental changes, and traffic loads often cause various localized abnormal waveforms in the monitoring time series, including isolated jumps, short-term spikes, continuous multi-point anomalies, positive and negative bipolar peaks, peak clusters, and short-term anomalies with low signal-to-noise ratios. These abnormal peaks may originate from transient structural responses or be caused by malfunctions in various parts of the acquisition system. Their appearance does not necessarily indicate structural damage; a comprehensive analysis combining waveform characteristics, responses of nearby sensors, and multi-source information is necessary. However, abnormal peaks can interfere with subsequent data statistics, status assessment, and damage identification. Therefore, accurate identification, determination of peak time, location range, and duration is crucial.

[0004] Existing detection methods mainly include amplitude thresholding, local statistical outlier detection, filtering, wavelet analysis, empirical mode decomposition (EMD), and machine learning. Amplitude thresholding is simple and fast, but the monitored signal is non-stationary, with significant differences in signal amplitude and noise levels across different channels, time periods, and operating conditions; a fixed threshold can easily lead to misjudgments. Local statistical outlier detection relies on neighborhood statistical characteristics, but its effectiveness in identifying continuous anomalies, low signal-to-noise ratio anomalies, and spike clusters is greatly affected by window and threshold settings. Filtering primarily focuses on noise reduction and cannot directly provide evidence of the start and end times of anomalies or their location. Empirical mode decomposition suffers from endpoint effects, mode aliasing, and parameter sensitivity when processing non-stationary signals. Machine learning methods, while capable of identifying complex patterns, have limitations such as training sample dependence, high computational cost, and insufficient interpretability. Wavelet analysis, for long-term, multi-channel, high-sampling-rate structural monitoring data, employs multi-scale wavelet analysis across a uniform preset scale range for the entire signal, resulting in a massive amount of coefficient calculation. In addition, multi-sensor consistency information helps to analyze the scope and source of the impact of abnormal events. However, when there are only a few candidate anomalies, continuously calculating the correlation and the proportion of synchronous anomalies for all channels and time periods will generate a large amount of computational overhead that is unrelated to the candidate anomalies. Summary of the Invention

[0005] To address the technical problems mentioned above, this invention provides a method and system for identifying and locating anomalies in time-series structural health monitoring signals. The method extracts candidate edges and forms candidate peak events based on differential sampling of adjacent sampling points and local robust statistics. It fuses multi-source width information that has undergone reliability checks to obtain candidate widths, and determines the local analysis window and wavelet scale set accordingly. Multi-scale wavelet analysis is performed within the local analysis window corresponding to the candidate peak events, estimating noise levels scale-by-scale and determining adaptive thresholds. A two-dimensional cross-scale response region is formed based on the adjacency relationship of the over-threshold response in both time and scale directions. The peak time is determined based on the time aggregation response within the response region, and the time projection of the response region is fused with the candidate edge interval to determine the location interval. This invention can balance the accuracy, computational efficiency, adaptability, and interpretability of anomaly peak identification in long-term, multi-channel structural health monitoring data, reducing false detections, missed detections, interval offsets, and computational burden caused by fixed thresholds, fixed scales, and global wavelet scanning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for identifying and locating anomalies in timing signals during structural health monitoring.

[0007] A method for identifying and locating anomalies in time-series signals during structural health monitoring, comprising: Acquire time-series signals from at least one structural monitoring sensor channel, calculate the difference between adjacent sampling points, construct a differential discriminant using the robust centrality and robust differential scale of the local statistical window, extract candidate edges of abnormal spikes, associate candidate edges of abnormal spikes in the abnormal entry and exit processes, and extract candidate spike events. From the event span determined by the candidate edges of paired abnormal peaks, the local waveform half-peak width, the local energy accumulation span, and the span between the significant difference positions on both sides of the local peak, at least two width information that have passed the reliability check are selected and robustly fused to obtain the candidate width, thereby determining the local analysis window and the wavelet scale set. In response to candidate spike events, for the local analysis window corresponding to the candidate spike event, multi-scale wavelet coefficients within the local analysis window are obtained according to the wavelet scale set. The noise level of the multi-scale wavelet coefficients is estimated scale by scale and an adaptive threshold is determined. Anomalies exceeding the corresponding adaptive threshold are obtained, and binary labels and overthreshold intensities are constructed. The time-scale grid points that meet the binary label conditions are connected according to the time and scale directions to obtain a two-dimensional cross-scale response region. The overthreshold intensities within the response region are aggregated to obtain the time aggregated response, so as to determine the peak time of the anomaly spike and thus determine the location interval.

[0008] Further, the method of extracting candidate edges of abnormal spikes to correlate candidate edges of abnormal entry and exit processes, and extracting candidate spike events, includes: The sampling positions where the differential discrimination value exceeds the candidate edge discrimination threshold are identified as abnormal spike candidate edges; The polarity of the candidate edges of abnormal spikes is determined based on the difference between adjacent sampling points and the robust center quantity. Abnormal spike candidate edges with a time interval not greater than a preset merging interval and the same polarity are merged into an abnormal spike candidate edge group; the abnormal spike candidate edge groups representing the abnormal entry process and the abnormal exit process have opposite polarities, and the time interval between them is between a preset minimum event span and a preset maximum event span; and according to at least one of amplitude matching relationship and local baseline change, the abnormal spike candidate edge groups representing the abnormal entry process and the abnormal exit process are associated to form candidate spike events; The amplitude matching relationship is as follows: determine the absolute amplitude between the candidate edge group of abnormal peaks representing the abnormal entry process and the candidate edge group of abnormal peaks representing the abnormal exit process. If the absolute amplitude meets the set conditions, the amplitude is considered to be matched. The local baseline change is defined as follows: when the normalized baseline change does not exceed a preset threshold, it is determined that the local baseline has been restored after the anomaly occurred.

[0009] Furthermore, the local waveform half-peak width is: a continuous half-amplitude support interval including the position of the local absolute amplitude peak; the local energy accumulation span is: the interval where the cumulative square deviation relative to the local robust baseline reaches 10% to 90% of the total energy; the span between the significant difference positions on both sides of the local peak is: the interval on both sides of the local absolute amplitude peak that satisfies the significant difference condition. The reliability check includes: determining whether the corresponding width information is a finite positive value, whether it is within a preset allowable range, and whether the corresponding half-width support, energy accumulation interval, or significant difference position completely falls within the initial width estimation range; discarding width information that fails the reliability check; The step of selecting at least two width information values ​​that have passed reliability checks and performing robust fusion to obtain candidate widths includes: setting a width consistency tolerance coefficient. γ >1. When there are at least three types of width information involved in the consistency screening, let m W =median W j Retaining 1 / γ ≤ W j / m W ≤ γ The width information; when there are exactly two types of width information involved in the consistency screening, if max(W 1, W 2) / min( W 1, W 2)≤ γ If two width information values ​​are retained, then two width information values ​​are retained; otherwise, it is determined to be a width evidence conflict, the preset width is adopted, and a width back-off flag is output. If at least two width information values ​​are retained after consistency screening, median fusion is performed; if fewer than two width information values ​​are retained, the preset width is adopted, and a width back-off flag is output. m W For those participating in the screening W j the median of W 1 and W 2 represents the two widths when there are only two reliable widths of evidence.

[0010] Furthermore, the local analysis window includes a window center, a local analysis window length, and a local analysis window width. The window center is the time center of the candidate spike event, the middle position of the candidate edge group of abnormal spikes representing the abnormal entry and exit process, or the extreme position of the local waveform corresponding to the candidate spike event. The local analysis window length is constrained by the candidate width, the minimum window length, and the maximum window length. The local analysis window width is determined by the window center and the local analysis window length.

[0011] Further, determining the adaptive threshold includes: At each scale, the median of the signed wavelet coefficients within the entire local analysis window is used as the robust central benchmark for noise scale estimation, and the noise scale is estimated by the absolute deviation of the wavelet coefficients from the median of the robust central benchmark. Based on the noise scale, the scale-by-scale normalized amplitude response is determined. From the scale-by-scale normalized amplitude response samples used for threshold estimation, the samples with the highest preset proportion of values ​​are excluded, and the adaptive threshold is determined based on the empirical tail quantile or the false alarm probability of the remaining samples.

[0012] Further, the method involves acquiring abnormal responses exceeding a corresponding adaptive threshold and constructing binary labels and overthreshold intensities; performing connectivity analysis on time-scale grid points satisfying the binary labeling conditions along the time and scale directions to obtain a two-dimensional cross-scale response region; and aggregating the overthreshold intensities within the response region to obtain a time-aggregated response to determine the peak time of the abnormal spike, thereby determining the location interval; the method includes: Binary markers and overthreshold intensities are constructed based on the scale-normalized amplitude response and adaptive threshold. Within the local analysis window corresponding to the current candidate peak event, markers with a value of 1 are aggregated to form one or more two-dimensional connected regions according to the 8-neighborhood relationship of temporal and scale adjacency. The 8-neighborhood relationship means that the absolute value of the scale index difference between any two markers with a value of 1 is not greater than 1, the absolute value of the temporal index difference is not greater than 1, and the two index differences are not both 0. make I k For the first k The projection of a two-dimensional connected region onto the time axis I e The time interval of the candidate abnormal spike event; will satisfy I k ∩ I e ≠ The two-dimensional connected region is determined as the two-dimensional cross-scale response region. When there are multiple two-dimensional cross-scale response regions, the two-dimensional cross-scale response region associated with the current candidate peak event is determined based on at least one of the following: the degree of overlap between the time projection of each two-dimensional cross-scale response region and the time interval of the candidate peak event, the sum of the overthreshold intensity of the region, and the time distance between the time center of the region and the center of the candidate peak event. The overthreshold intensity of each scale in the two-dimensional cross-scale response region is weighted and summed to obtain the time aggregation response, and the time corresponding to the maximum value of the time aggregation response is taken as the peak time. The projection of the two-dimensional cross-scale response region on the time axis is determined as the first positioning interval, and the first positioning interval is combined with the second positioning interval determined by the candidate entry edge and exit edge by taking the envelope or fusing according to the confidence level to obtain the positioning interval of the candidate peak event.

[0013] A second aspect of the present invention provides a structural health monitoring timing signal spike anomaly identification and location system.

[0014] A structural health monitoring time-series signal spike anomaly identification and localization system includes: The signal acquisition and candidate event generation module is configured to: acquire time series signals from at least one structural monitoring sensor channel, calculate the difference between adjacent sampling points, construct a differential discriminant using the robust centrality and robust differential scale of the local statistical window, extract candidate edges of abnormal peaks, associate candidate edges of abnormal peaks in the abnormal entry and exit process, and extract candidate peak events. The candidate width estimation module is configured to: select at least two width information that have passed the reliability check from the event span determined by the candidate edges of the paired abnormal peaks, the local waveform half-peak width, the local energy accumulation span, and the span between the significant difference positions on both sides of the local peak, and perform robust fusion to obtain the candidate width, thereby determining the local analysis window and the wavelet scale set; The cross-scale aggregation localization module is configured to: respond to candidate spike events, obtain multi-scale wavelet coefficients within the local analysis window corresponding to the candidate spike event according to the wavelet scale set, estimate the noise level of the multi-scale wavelet coefficients on a scale-by-scale basis and determine an adaptive threshold, acquire abnormal responses exceeding the corresponding adaptive threshold, and construct binary labels and overthreshold intensities; perform connectivity analysis on the time-scale grid points that meet the binary label conditions according to the time and scale directions to obtain a two-dimensional cross-scale response region, and aggregate the overthreshold intensities within the response region to obtain a time aggregation response, thereby determining the peak time of the abnormal spike and thus determining the localization interval.

[0015] A third aspect of the present invention provides a computer device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the structural health monitoring timing signal spike anomaly identification and localization method as described in the first aspect above.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and to execute the steps of the structural health monitoring timing signal spike anomaly identification and localization method as described in the first aspect above.

[0017] The fifth aspect of the present invention provides a computer program product or computer program.

[0018] This invention provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the structural health monitoring timing signal spike anomaly identification and localization method as described in the first aspect above.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention employs local robust difference analysis to pre-screen candidate data from all time periods of monitoring data. The required calculations mainly involve the difference between adjacent sampling points and local robust statistics. The calculation process is simple and suitable for processing long-term, multi-sampling-point bridge monitoring time series.

[0020] In response to the generated candidate spike events, multi-scale wavelet analysis for obtaining abnormal responses is performed within the corresponding local analysis window, so that the wavelet calculation range is concentrated in the vicinity of the candidate events; in a preferred embodiment, the multi-scale wavelet analysis is not performed outside the local analysis window, thereby avoiding including the entire monitoring period in the high-cost wavelet scan.

[0021] This invention estimates the candidate width based on the local morphology of the candidate peaks, and uses the candidate widths to jointly constrain the local analysis window and wavelet scale set, so that the analysis parameters can change with the continuous width of the abnormal peaks. It is applicable to the identification and localization of isolated jump points, short-term spikes, anomalies of multiple consecutive sampling points, bipolar peaks, peak clusters, and short-term anomalies with low signal-to-noise ratio.

[0022] Based on the actual wavelet coefficient distribution at each wavelet scale, this invention estimates the robust central quantity and noise scale on a scale-by-scale basis, and determines the adaptive threshold on a scale-by-scale basis. It can adapt to noise differences and non-stationary backgrounds under different sensor channels, different monitoring periods, and different wavelet scales.

[0023] This invention forms a two-dimensional cross-scale response region based on the adjacency relationship of the abnormal response in the time and scale directions, and determines the peak time and positioning interval through time-scale binary labeling, overthreshold intensity, two-dimensional connected region and time aggregation response. This helps to reduce the dependence of peak positioning on local fluctuations of a single scale and provides verifiable and interpretable cross-scale evidence for the positioning interval.

[0024] By fusing at least two width information from the event span, local waveform half-peak width, local energy cumulative span from 10% to 90%, and span between significant difference positions on both sides of the local peak to determine the candidate edge, and excluding width information that cannot be reliably obtained, it helps to reduce the impact of distortion of single width information on the determination results of local window and scale range.

[0025] This invention can output structured results such as peak time, location interval, duration, shape, candidate width, dominant scale, scale range, event confidence, and multi-sensor consistency index of abnormal spikes, which facilitate manual review, data quality recording, and bridge monitoring and maintenance analysis. Attached Figure Description

[0026] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0027] Figure 1 This is a flowchart illustrating the method for identifying and locating anomalies in timing signals during structural health monitoring, as shown in an embodiment of the present invention. Figure 2This is a flowchart of another embodiment of the structural health monitoring time-series signal spike anomaly identification and localization method shown in this invention; Figure 3 This is a schematic diagram of differential candidate edge extraction shown in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating at least two types of width information and robust fusion in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the determination of the local analysis window and wavelet scale set in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating local multi-scale wavelet analysis and scale-by-scale adaptive thresholding in an embodiment of the present invention; Figure 7 This is the two-dimensional multi-scale connected region Ω and aggregate response shown in the embodiments of the present invention. G ( t (and a schematic diagram of the positioning range); Figure 8 This is a schematic diagram of the candidate-triggered local multi-scale analysis and two-dimensional cross-scale localization process of event number 1 shown in an embodiment of the present invention; Figure 9 This is a schematic diagram illustrating long-time sequence overlapping block segmentation and cross-block event deduplication in an embodiment of the present invention; Figure 10 This is a structural diagram of a structural health monitoring timing signal spike anomaly identification and location system according to an embodiment of the present invention; Figure 11 This is a structural diagram of a computer device shown in an embodiment of the present invention. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0030] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0031] As described in the background section, existing technologies need to address the following issues: how to reduce unnecessary multi-scale wavelet calculations in long-term, multi-channel bridge monitoring data; how to adapt the local analysis window, wavelet scale range, and discrimination threshold to the width of candidate peaks and local noise background; how to utilize the adjacency relationships of anomalous responses in the time and scale directions to form interpretable and verifiable peak times and location intervals; and how to call upon information from neighboring sensors for auxiliary analysis during the local time period corresponding to candidate anomalies. To solve at least one of the above technical problems, this invention provides a method and system for identifying and locating peak anomalies in structural health monitoring time-series signals. The following detailed description of the invention's solution, in conjunction with several embodiments, further illustrates this.

[0032] Figure 1 This is a flowchart illustrating a method for identifying and locating anomalies in timing signals during structural health monitoring, as shown in an embodiment of the present invention; see reference. Figure 1 The method includes: Acquire time-series signals from at least one structural monitoring sensor channel, calculate the difference between adjacent sampling points, construct a differential discriminant using the robust centrality and robust differential scale of the local statistical window, extract candidate edges of abnormal spikes, associate candidate edges of abnormal spikes in the abnormal entry and exit processes, and extract candidate spike events. From the event span determined by the candidate edges of paired abnormal peaks, the local waveform half-peak width, the local energy accumulation span, and the span between the significant difference positions on both sides of the local peak, at least two width information that have passed the reliability check are selected and robustly fused to obtain the candidate width, thereby determining the local analysis window and the wavelet scale set. In response to candidate spike events, for the local analysis window corresponding to the candidate spike event, multi-scale wavelet coefficients within the local analysis window are obtained according to the wavelet scale set. The noise level of the multi-scale wavelet coefficients is estimated scale by scale and an adaptive threshold is determined. Anomalies exceeding the corresponding adaptive threshold are obtained, and binary labels and overthreshold intensities are constructed. The time-scale grid points that meet the binary label conditions are connected according to the time and scale directions to obtain a two-dimensional cross-scale response region. The overthreshold intensities within the response region are aggregated to obtain the time aggregated response, so as to determine the peak time of the anomaly spike and thus determine the location interval.

[0033] This invention first extracts candidate edges of anomalous spikes using local robust differential statistics, and forms candidate spike events based on the polarity, temporal relationship, amplitude relationship, and local baseline recovery of the candidate edges. Then, it robustly fuses at least two width information to obtain candidate widths, and adaptively determines the local analysis window and wavelet scale set accordingly. Within the local analysis window, it performs multi-scale wavelet analysis, scale-by-scale noise estimation, and adaptive threshold determination. Finally, it determines the peak time of the anomalous spike and the final location interval through cross-scale connectivity analysis and scale aggregation response on a time-scale two-dimensional grid.

[0034] The following is a detailed description of the method for identifying and locating spike anomalies in the time-series signals of structural health monitoring, such as... Figure 2 As shown, it includes the following steps: S1: Acquire the time series signal of structural monitoring.

[0035] Acquire time-series signals from at least one structural monitoring sensor channel. The signals can be strain, displacement, acceleration, tilt angle, temperature, wind speed, or other monitored quantities.

[0036] by N Taking a single-channel signal consisting of 10 sampling points as an example, it is denoted as:

[0037] in, x ( n ) indicates the first n The monitoring values ​​of each sampling point.

[0038] S2: Difference and Local Robust Statistics for Extracting Candidate Edges.

[0039] like Figure 3 As shown, calculate the first-order difference between adjacent sampling points:

[0040] For the i At each sampling location, within its corresponding local statistical window, the robust central quantity is determined based on the background difference samples. and robust difference scale And a positive robust difference scale was obtained after validity checks. Based on this, a local difference discriminant is constructed:

[0041] In the formula, ε d As a preset positive number, and They have the same dimensions to avoid numerical instability caused by excessively small robust difference scales; It is a dimensionless difference discriminant.

[0042] The robust centrality measure can be the differencing median, the truncated mean, or other central statistics that are insensitive to outliers. Taking the median as an example:

[0043] In the formula, r for B i The location index of the inner difference sample; B i In order to be with the first i The set of valid background difference sample indices in the local statistical window corresponding to each sampling location; d ( r ) is the first r The difference samples; median{·} represents taking the median of the samples in the set.

[0044] The robust differential scale can be the absolute deviation of the median, the interquartile range converted value, the truncated standard deviation, or a combination thereof. Taking the absolute deviation of the median as an example:

[0045] Wherein, 0.6745 is an approximate value of the consistency correction coefficient for the absolute deviation of the median under normal noise conditions; when using other noise models, the corresponding consistency correction coefficient can also be used.

[0046] When the estimated When the value is non-numerical, infinite, zero, negative, or below the preset minimum scale, a backtracking can be performed using the interquartile range conversion value, sample standard deviation, the scale corresponding to the sensor resolution, or the preset lower limit of the positive scale to obtain an effective positive scale. .

[0047] When the differential discriminant exceeds the candidate edge discrimination threshold T d At that time, the first i Each sampling location was determined as a candidate edge:

[0048] The polarity of candidate edges can be determined according to the following relationship:

[0049] In the formula, sgn(·) is the sign function; - The signed bias of the current difference relative to the local robust center; i , and The meaning is the same as described above. When the deviation is positive, negative, or zero, Take +1 respectively, 1 or 0, where It is not considered as a positive or negative candidate edge.

[0050] Indicates the positive candidate edge. This represents a negative candidate edge. All candidate edges constitute the candidate edge set:

[0051] By using the aforementioned local robust differential pre-screening, time locations that may be associated with anomalous peaks can be prioritized without performing a full-scale wavelet scan on the complete time series.

[0052] S3: Candidate edge grouping and combination form candidate spike events.

[0053] Candidate edges that are temporally adjacent, have the same polarity, and meet the preset interval conditions within a group are merged into a candidate edge group.

[0054] The preset interval within a group can be represented by the number of sampling points. N merge It can also be expressed as time. T merge Indicates; when given in time, the sampling frequency is... f s and T merge Round the product up to the nearest integer to convert it to the number of sampling points. N merge . N merge It can be determined based on the minimum expected peak width, sensor sampling frequency, and allowable intra-group edge discontinuities; the sampling point interval between adjacent candidate edges of the same polarity should not exceed [a certain value]. N merge When they are identified as belonging to the same candidate edge group, they are determined to be of similar origin.

[0055] For the entering and exiting edge groups to be associated, the signed group amplitude is determined by the sum of the original differences within the group, the edge amplitude with the largest absolute value, or the weighted sum of the edge amplitudes within the group, respectively. A in and A out .

[0056] Furthermore, the candidate edge groups representing the abnormal entry and exit processes must at least satisfy the following conditions: the exit edge group is located after the entry edge group, the two have opposite polarities, and the time interval is between the preset minimum and maximum event spans; and at least one of the amplitude matching condition and the local baseline recovery condition is met. When the above conditions are met, the candidate edge groups representing the abnormal entry and exit processes are associated to form candidate spike events.

[0057] For positive and negative spikes, the polarity order of entry and exit can be reversed, but they are still required to be of opposite polarities. For bipolar short-time anomalies containing multiple lobes that are either positive first and then negative or negative first and then positive, multiple candidate edge groups that meet the correlation conditions can be combined to form a candidate anomaly event based on the temporal relationship between adjacent candidate edge groups.

[0058] For the amplitude matching relationship between entering and exiting the edge group, the ratio of the smaller absolute amplitude to the larger absolute amplitude can be used. R A Judgment; when R A Not lower than the preset threshold ρ A The amplitude is considered to be matched when 0 < ρ A ≤1. Equivalently, the ratio of the larger absolute amplitude to the smaller absolute amplitude can also be limited to not exceeding a preset upper limit.

[0059] To evaluate the local baseline recovery, background segments were selected before entering the edge group and after exiting the edge group, and the background median was calculated. b before and b after The background segment length can be determined based on the initial event span between the current entry edge group and the exit edge group, the preset background length, or a combination of the two, without relying on the candidate width obtained in step S4, and avoids the current candidate event area and its preset protection area.

[0060] Let the robustness scale of background fluctuations be... s base After validity checks, a positive metric was obtained. base Then the normalized baseline change can be expressed as:

[0061] Where, Δ base Indicates the change in normalized baseline. b before This indicates the background median of the background segment selected before entering the edge group. b after This indicates the background median of the selected background segment after exiting the edge group.

[0062] When Δ base Not exceeding the preset threshold ρ base At that time, it was determined that the local baseline had been basically restored after the anomaly occurred.

[0063] Candidate spike events are used to limit the scope of interest for subsequent width estimation and local multi-scale analysis. The candidate events formed in this step are not directly used as the final anomaly identification and localization conclusions.

[0064] Determine the time interval of candidate events based on candidate entry edges and candidate exit edges. Ie =[ t enter , t exit ], and construct the initial width estimation range. W 0=[ t enter - Δ 0, t exit + Δ The intersection of 0] and the current record range, where Δ 0 represents the preset protection length. Initial width estimation range. W 0 is only used for width information calculation and integrity check in step S4, and does not depend on the candidate width obtained in step S4.

[0065] S4: Combine at least two width information to obtain candidate widths.

[0066] This embodiment selects at least two reliably obtainable types of width evidence from the following four categories for fusion: event span determined by paired candidate edges, local waveform half-peak width, cumulative span of local energy from 10% to 90%, and span between significant difference positions on both sides of a local peak. The widths are defined as follows:

[0067]

[0068]

[0069]

[0070] in, W edge The event span determined by paired candidate edges, W half The half-peak width of the local waveform. W energy For the local energy cumulative span of 10% to 90%, W diff The span between the significant difference positions on both sides of the local peak; t enter and t exit These represent the moments when pairs of candidates enter the edge and candidates exit the edge, respectively. t 0.5,L and t0.5,R These are the left and right endpoints of the half-amplitude continuous support interval containing the local absolute amplitude peak; t 10% and t 90% These represent the moments when the local energy accumulates to 10% and 90% of the total energy, respectively. t d,start and t d,end These are the significant difference positions selected on the left and right sides of the local peak according to preset rules. All four widths are obtained by subtracting the left endpoint from the corresponding right endpoint; when expressed as time, the dimension is time; when expressed as the number of sampling points, the dimension is the number of sampling points.

[0071] Within the initial width estimation range W Within 0, the local waveform amplitude is calculated using the absolute deviation relative to the local robust baseline, thus applying the same width calculation caliber to both positive and negative spikes. The left and right endpoints of the half-peak width are the two endpoints of a continuous half-amplitude support interval containing the peak position of the local absolute amplitude; the endpoints of the energy accumulation span are the positions where the accumulated square deviation relative to the local robust baseline reaches 10% and 90% of the total energy; the two endpoints of the significant difference span are the positions on both sides of the peak of the local absolute amplitude that satisfy the significant difference condition. When multiple positions satisfying the condition exist on the same side, the one with the largest absolute difference value is selected according to a preset rule. Bipolar spikes or spike clusters determine the local robust baseline, cumulative energy range, and left and right side positions using the overall envelope of the candidate event. Any required endpoint is missing or the corresponding support interval is touched. W When the boundary is 0, the corresponding width information is determined to be unreliable.

[0072] Before fusion, all width information is standardized to the same unit. For width information expressed in time, the time unit can be directly adopted; for width information expressed in the number of sampling points, the sampling frequency can be used. f s This is converted to the corresponding time width. Alternatively, the width information expressed in time can be converted to the number of sampling points before fusion. This method ensures that width information from different sources has a consistent dimension, providing a unified width benchmark for determining the subsequent local analysis window and wavelet scale set. In subsequent use, when the candidate width uses different units of representation than the local analysis window length or wavelet scale, the width can be determined based on the sampling frequency. f s Perform the corresponding conversion.

[0073] Width information is deemed unreliable if any of the following conditions are met: complete paired edges are not formed; the half-width, energy accumulation interval, or significant difference location is incomplete or cannot be within the initial width estimation range. WReliable determination within 0; width is non-numerical, infinite, zero, negative, or exceeds the allowable range; corresponding support interval touches W 0 boundary.

[0074] Discard width information that cannot be reliably obtained; let γ >1. When there are at least three types of width information involved in the consistency screening, let m W =median W j Retaining 1 / γ ≤ W j / m W ≤ γ The width information; when there are exactly two types of width information involved in the consistency screening, if max( W 1, W 2) / min( W 1, W 2)≤ γ If the width information is not less than two types, then two width information types are retained; otherwise, it is determined to be a width evidence conflict, and the preset width is adopted and a width rollback flag is output. If the width information retained after consistency screening is not less than two types, median fusion is performed; if it is less than two types, the preset width is adopted and a width rollback flag is output.

[0075] In the formula, γ For width consistency tolerance coefficient and γ >1; W j For the first j The item has passed reliability checks and has been standardized in its breadth of evidence. j For the width of the evidence sequence number; m W For those participating in the screening W j the median; W 1 and W 2 represents the two widths when there are only two reliable width pieces of evidence. All the width ratios mentioned above are ratios of quantities with the same dimension, and therefore are dimensionless.

[0076] like Figure 4 As shown, Figure 4 (a) shows the candidate local waveforms and four types of width evidence, the W edge , W half , W energy and W diff As Figure 4 The inputs for item-by-item reliability checks and consistency screening in (b) of the document; Figure 4(b) First, remove invalid evidence and then proceed according to... γ Perform consistency screening, then perform median robust fusion or preset width backoff based on the number of reliable evidence, and output candidate widths. And the width_fallback_used flag. Each piece of evidence comes from a different local geometry or energy definition, and the same event span is not counted repeatedly in both types of evidence.

[0077] S5: Determine the local analysis window and wavelet scale set based on the candidate width.

[0078] Based on the candidate spike events formed in step S3 and the candidate widths obtained in step S4, the local analysis window and wavelet scale set are determined, such as... Figure 5 (a) Figure 5 As shown in (b) of the diagram.

[0079] Center position of local analysis window t c The window center can be determined by the time center of the candidate spike event, the midpoint between the candidate entry edge and exit edge, or the extreme position of the local waveform corresponding to the candidate spike event. The extreme position can be the position with the largest absolute deviation from the local baseline within the range of the candidate event. Among these, the midpoint between the candidate entry edge and exit edge is preferred to determine the window center; when the paired edge information is incomplete, the extreme position of the local waveform corresponding to the candidate spike event can be used.

[0080] The local analysis window length is determined based on the candidate width and the preset minimum and maximum window length limits:

[0081] in, L w This is the length of the local analysis window. L min and L max These are the preset minimum window length and maximum window length, respectively. β This is the window expansion factor. The candidate width.

[0082] According to the window center t c and window length L w Determine the width of the local analysis window W L :

[0083] in, t L and tR These are the left and right boundaries of the local analysis window, respectively, and based on... t c and L w Determine; when the determined window exceeds the original signal boundary, truncate the corresponding boundary to the effective sampling range.

[0084] Further, based on the candidate width obtained in step S4 Determine the discrete wavelet scale set A and its upper and lower bounds:

[0085] in:

[0086]

[0087] In the formula, a For wavelet scale, The preset maximum allowable size, κ This is the scale range adjustment coefficient; and These are the minimum and maximum scales of the scale set, respectively; floor(·) and ceil(·) represent rounding down and rounding up, respectively.

[0088] Therefore, the length of the local analysis window and the wavelet scale range are adaptively determined based on the candidate width of the candidate peak event, so that the subsequent wavelet analysis is focused on the vicinity of the candidate event and the scale range adopted is matched with the duration of the candidate peak.

[0089] S6: Candidate triggered local multi-scale wavelet analysis, scale-by-scale noise estimation, and determination of adaptive threshold.

[0090] like Figure 6 As shown, in response to candidate spike events, Figure 6 (a) in the diagram is a schematic of a local analysis window. Figure 6 (b) in the table represents the signed wavelet coefficients. C ( a , t ) Schematic diagram, Figure 6 (c) in the figure is the scale-normalized amplitude response. Schematic diagram; for the local analysis window determined in step S5 W L According to wavelet scale set A get W L The multi-scale wavelet coefficients within the set. This embodiment uses the scale set... AThe constructed zero-mean normalized wavelet kernel is convolved with the local detrended signal to obtain wavelet coefficients at the selected discrete scale. This method is a numerical implementation of continuous wavelet transform on discrete sampled signals, which does not depend on specific software functions or toolboxes; the "local multi-scale wavelet analysis" described in this paper includes the above wavelet kernel convolution implementation and its equivalent continuous wavelet transform implementation.

[0091] The continuous wavelet transform can theoretically be expressed as:

[0092] in, x L ( t () is the local analysis window W L Signals after detrending or median removal processing according to the Inner Canon of Medicine. a For wavelet scale, b The position is the wavelet translation. ψ ( ) is the mother wavelet function, ψ ( ) indicates its complex conjugate. C ( a , b ) represents the wavelet coefficients at the corresponding scale and location.

[0093] For discretely sampled structural health monitoring signals, the above integration is implemented using discrete convolution or an equivalent discrete summation method, and only on the scale set. A Included scales and local analysis windows W L The calculation is performed at the included sampling locations.

[0094] At every scale Above, the median of the signed wavelet coefficients within the entire local analysis window. As a robust central benchmark for noise scaling estimation:

[0095] The noise scale is estimated using the absolute deviation of the wavelet coefficients relative to the robust central benchmark:

[0096] The value of 0.6745 in the formula is an approximate value of the normal consistency correction coefficient; when using the normal noise assumption, it can be taken as about 0.6745, or other consistency correction coefficients that match the target noise distribution can be used.

[0097] when σ aIf the value is non-numerical, infinite, zero, negative, or below the preset minimum noise scale, the estimate is deemed invalid. The backoff scale can be the normal consistency conversion value obtained by dividing the sample standard deviation or interquartile range of the wavelet coefficients within the local analysis window by approximately 1.349, or a positive number determined based on sensor resolution, numerical accuracy, or a preset noise lower limit. After applying the backoff scale, it is still recorded as... σ a .

[0098] Because the input local signal has been detrended or median-removed and the wavelet function is constructed using zero mean, a robust central benchmark is established. m a Used to calculate noise scale σ a The abnormal response is normalized using the absolute amplitude of the wavelet coefficients instead of being repeatedly subtracted. m a The scale-normalized amplitude response is:

[0099] In the formula, ε w This is a preset positive number in the wavelet coefficient normalization, and... σ a They have the same dimensions to avoid numerical instability caused by excessively small noise scales; z w ( a , t ) represents the dimensionless, scale-normalized amplitude response.

[0100] To reduce the influence of the candidate peaks themselves on the empirical tail fractions, at each scale a Above, samples with the highest values ​​are excluded from the scale-normalized amplitude response samples used for threshold estimation, representing a predetermined proportion. As an example, the scale-wise... z w ( a,t Samples not higher than the 95th percentile are retained to form the threshold estimation sample set:

[0101] in, Q q ( ) represents the sample q Quantiles. This exclusion operation applies only to the threshold estimation sample set and does not remove the original scale-wise responses used for anomaly detection and localization.

[0102] Based on the truncated background response samples, an adaptive threshold is determined at each scale. T a :

[0103] in, p The empirical tail probability parameter is set for the truncated background response sample set. Indicates experience 1 p-quantiles, T min This is a preset minimum threshold, used to prevent the threshold from dropping abnormally when the local background response is too low.

[0104] S7: Two-dimensional cross-scale connectivity aggregation and localization.

[0105] like Figure 7 As shown, the scale-by-scale normalized amplitude response obtained in step S6 z w ( a,t and the corresponding adaptive threshold T a Constructing binary labels M ( a , t and overthreshold intensity S ( a , t ):

[0106]

[0107] Among them, 1[ ] is the indicator function, [ ] + This indicates taking the non-negative part. When... z w ( a,t )> T a hour, M ( a , t )=1, and S ( a , t )>0; otherwise M ( a , t )=0, and S ( a , t )=0.

[0108] Within the local analysis window corresponding to the current candidate spike event, the discrete-time location is compared with the ordered wavelet scale set determined in step S5:

[0109] Construct a time-scale two-dimensional grid and satisfy... M (a j , t n Grid points with )=1 are analyzed for connectivity based on their 8-neighborhood relationships in both the time and scale directions to obtain two-dimensional connected regions, such as... Figure 7 As shown in (a):

[0110] Among them, CC8 ( ) indicates that according to the 8-neighbor relationship, the condition is satisfied. M ( a j , t n The time-scale grid points with a time scale of 1 are used to divide the region into connected components, Ω. k Indicates the first k A two-dimensional connected region, K This represents the number of connected components obtained. Furthermore, a j scale set A The first in j One scale, t n For the first in the local analysis window n Discrete time positions j and n These are the scale number and the time number, respectively; M ( a j , t n ) represents the binary overthreshold label for the corresponding time-scale grid point. M ( a j , t n )=1 indicates that the scale-wise normalized amplitude response at this point exceeds the corresponding scale threshold; k This represents the sequence number of the connected region. k =1,…, K .

[0111] For two different overthreshold grid points ( a j , t n )and( a j' , t n' When the following conditions are met: and At that time, the two are determined to be 8-neighbors. Adjacency relationships in the scale direction are determined according to the scale set. AThe order of the various scales is determined, but the difference between the actual scale values ​​of adjacent scales is not required to be equal.

[0112] like Figure 7 As shown in (b), let the time interval determined by the candidate entry edge and the candidate exit edge of the current candidate spike event be:

[0113] In the formula, I e This represents the candidate time interval for the current candidate peak event; t e,L and t e,R These are the candidate entry edge time and candidate exit edge time for the event, respectively. t e,L ≤ t e,R All three use the same time unit as the original timeline.

[0114] For the k A two-dimensional connected region Ω k Its projection onto the time axis is denoted as:

[0115] In the formula, Proj t (·) indicates that a two-dimensional connected region is projected onto the time axis; I k Ω is a two-dimensional connected region k The projection interval on the time axis; t k,L and t k,R These are the left and right endpoints of the projection interval, respectively.

[0116] in,

[0117] According to the time projection I k Time interval of the current candidate peak event I e The positional relationship between them is used to determine the region associated with the current candidate spike event from the two-dimensional connected region. When At that time, the corresponding two-dimensional connected region Ω can be... kCandidate association regions that meet the association criteria are identified; further, the time distance between the time center of the two-dimensional connected region and the center of the candidate peak event can be limited to not exceeding a preset association threshold. The time center of the two-dimensional connected region can be taken as the midpoint of its time projection interval, and the center of the candidate peak event can be taken as the midpoint of the time interval of the candidate peak event.

[0118] The corresponding center time distance is:

[0119] When multiple candidate associated regions satisfying the aforementioned association conditions exist, selection can be made based on at least one or a combination of the following: the degree of overlap between the projection of each candidate associated region on the time axis and the time interval of the candidate peak event; the sum of the region's overthreshold intensity; and the time distance between the region's time center and the candidate peak event center. The degree of overlap can be characterized by the intersection length or overlap ratio of two time intervals. Preferably, a two-dimensional connected region with a large degree of overlap, a large sum of the region's overthreshold intensity, and / or a small center time distance is selected, and the finally determined associated region is denoted as Ω. Wherein, the... k The sum of the superthreshold intensities of each two-dimensional connected region can be expressed as:

[0120] The time-aggregated response is obtained by scale-weighted aggregation of the overthreshold intensities at different scales within the associated region Ω. G ( t ):

[0121] Among them, weight ρ a ≥0 represents the scale weight. All scales can use the same weight, or different weights can be set according to the response confidence of different scales; when using normalized weights, we can let:

[0122] according to G ( t Determine the location of the aggregate response peak of candidate spike events. t p :

[0123] in, I Ω Projection of the associated region Ω onto the time axis:

[0124] and

[0125] when G ( t When multiple locations have the same maximum value, the location with the smallest time distance from the center of the candidate peak event can be preferentially selected as the peak moment. t p .

[0126] Projecting the time I Ω This serves as the first positioning interval, and the time interval for candidate spike events is determined by the candidate entry edge and the candidate exit edge. I e As the second positioning interval.

[0127] To reduce the possibility that the first positioning interval truncates the full time range of the peak when the two-dimensional overthreshold response is sparse, it is preferable to take the minimum envelope of the first and second positioning intervals to obtain the final positioning interval:

[0128] In another embodiment, the endpoints of the first and second positioning intervals can be selected or merged based on their respective positioning reliability to determine the final positioning interval, such as... Figure 7 As shown in (c) in the figure.

[0129] When a two-dimensional connected region associated with the current candidate spike event is not obtained, a backtracking approach can be used to construct a backtracking response within the current local analysis window based on the maximum response or weighted aggregate response of each scale's overthreshold intensity. For example:

[0130] Alternatively, a weighted sum of responses at each scale can be used:

[0131] Based on the rollback response within the candidate spike event time interval I e The location of the maximum value within determines the peak time. t p ,Right now:

[0132] When multiple locations with the same maximum value exist in the above-mentioned aggregate response or backoff response, the location with the smaller time distance from the center of the candidate spike event can be preferably selected as the peak moment. t p .

[0133] At this point, the candidate peak event time interval is used. Ie As the final positioning range I loc The output results are marked with the backtracking method used. The backtracking method is used when no effective associated two-dimensional connected regions are obtained, without changing the structure of the two-dimensional cross-scale connected aggregation positioning as the preferred implementation.

[0134] As an optional implementation, in the aforementioned main implementation, based on the entire local analysis window W L Median absolute deviation (MAD) noise estimation is performed on the wavelet coefficients at each scale. For scenarios where the candidate core region has strong energy and may significantly affect the local noise estimation, a core region can be selectively constructed. W C Nature Reserve W G and background outside the core area B W :

[0135]

[0136]

[0137]

[0138]

[0139] in, t C1 and t C2 These represent the start and end times of the candidate core region, respectively. Δ G This refers to the protection time span by which the protected area expands to both sides relative to the candidate core area. Protected Area W G For the expanded time interval and local analysis window W L The intersection, thus W G Limited to a local analysis window W L Inside; background outside the core area B W From the local analysis window W L Remove protected areas W G The subsequent sample composition.

[0140] For each scale, the background outside the core area is used. B WThe corresponding wavelet coefficients are used to calculate the robust center at this scale. and noise scale This allows for the estimation of noise by scale under background conditions outside the core area.

[0141] when B W When there are sufficient effective samples, a scale-normalized amplitude response is constructed based on the background noise scale outside the core region. In the formula, ε w This has the same meaning as in the aforementioned wavelet coefficient normalization; This represents the scale-normalized amplitude response under background conditions outside the core area. t ∈ B W of The sample with the highest preset proportion of values ​​is excluded to obtain the background threshold estimation sample outside the core area. and in accordance with Determine the adaptive threshold for the background outside the core area, and then use Determine the location of the abnormal response. When B W Insufficient number of valid samples If invalid or empirical quantiles cannot be reliably determined, the algorithm reverts to the corresponding scale noise scale and adaptive threshold determined based on the entire local analysis window; if it still cannot be effectively determined, a preset positive noise scale and a preset minimum threshold are used. Background estimation outside the core area is used as an optional enhancement method for the aforementioned scale-by-scale noise estimation when the candidate core area may have a significant impact on the local noise estimation.

[0142] As an optional implementation, candidate-triggered multi-sensor consistency is employed. In response to a formed candidate spike event, channel data from channels grouped with or spatially adjacent to the target channel are read within a local analysis window corresponding to the candidate spike event. Let the total number of neighboring channels be... M This excludes the target channel where the current candidate spike event is located; the number of neighboring channels that exhibit differential out-of-threshold responses within the preset time tolerance range corresponding to the candidate spike event is... n hit The proportion of abnormal synchronization channels is:

[0143] in, R sync This indicates a consistency index for multi-sensor synchronization anomalies. When... M When =0, multi-sensor consistency correction is not performed.

[0144] Differential evidence can be used P d Wavelet evidenceP w and consistency indicators R sync The event confidence scores are merged into event confidence scores, and these scores are output together with the candidate spike events. The event confidence scores are used for sorting, grading, or manual review, but are not a necessary condition for automatically excluding candidate events. The event confidence scores are determined according to the following formula:

[0145] In the formula, This represents the weighting coefficient. The event confidence level is used for ranking, classifying, or manually verifying candidate spike events. Multi-sensor consistency results are not a necessary condition for automatically excluding candidate spike events.

[0146] In response to candidate spike events formed in the target sensor channel, it is only necessary to read the same group or spatially adjacent sensor channels and calculate the consistency index within the corresponding local analysis window. This can provide auxiliary evidence for the confidence output, sorting, grading and manual review of candidate spike events while controlling the multi-channel processing range.

[0147] The following description uses the structural health monitoring of a bridge as an example to illustrate the method for identifying and locating anomalies in the timing signals of structural health monitoring as described in this invention.

[0148] (1) Data acquisition From one month of structural health monitoring data for a bridge, raw MAT files for three days, one hour per day, were pre-selected for validation. The selected data included monitoring data at both high and low sampling frequencies. The high sampling frequency data had a sampling rate of 51.2 Hz, corresponding to 342 channels, and the monitored quantities mainly included acceleration and dynamic strain. The low sampling frequency data had a sampling rate of 2.56 Hz, corresponding to 216 channels, and the monitored quantities mainly included displacement, rotation angle, and wind. The aforementioned analysis file contains time series data for all 558 channels.

[0149] A 60-second block-based robust differential pre-screening was performed on the data from all the aforementioned channels, and the robust normalization results of the differences between adjacent sampling points in each channel were calculated. To control the computational load and output scale of this experimental verification, the channels were sorted according to the maximum absolute robust normalized difference value obtained in the pre-screening stage, and the 45 channels with the highest scores were selected to perform candidate edge formation, candidate spike event combination, and local two-dimensional cross-scale localization.

[0150] The channel selection described above is only used to determine the verification range of the existing complete positioning results in this embodiment and does not constitute a criterion for determining candidate spike events. Whether other channels not included in the above verification range have candidate spike events that meet the determination criteria is not determined by the complete positioning results of this instance. The program reads the data matrix in the original MAT file and maps the matrix column numbers to channel names, units, and location descriptions according to the verified channel configuration information; only the original file is read during the analysis process.

[0151] (2) Specific processing procedure For each channel entering the complete positioning process, the median and MAD scale of the difference between adjacent sampling points are first calculated according to the 60 s data block to obtain the robust standardized difference; the sampling positions where the absolute standardized difference exceeds the preset threshold are combined as candidate edges, and then candidate spike events are formed according to the polarity, time interval and jump amplitude matching relationship of the candidate edges.

[0152] For each analysis channel, candidate spike events are sorted in descending order of differential peak evidence. In this practical run, complete local localization is performed on the top 10 candidate events. For candidate spike events entering the local localization process, the event span, half-width, and cumulative energy width (10% to 90%) are calculated, and the local analysis window and scale set are determined accordingly. Robust noise estimation and empirical tail quantile thresholds are used at each scale to construct time-scale binary labels. M ( a , t ) and continuous overthreshold intensity S ( a , t Subsequently, a two-dimensional connected region Ω associated with the current candidate is obtained according to the time-scale 8-neighborhood, and a aggregated response is constructed based on the continuous superthreshold intensity at each scale within Ω. G ( t ), to determine the peak time and positioning range.

[0153] If adjacent differential candidates converge to the same peak value or form highly overlapping positioning intervals, they are merged before the final output. For events where a valid two-dimensional connected region Ω cannot be formed after candidate gating, a safe backoff is performed using the maximum response at a one-dimensional scale, and the positioning mode is marked as connected2d_fallback_max. Width reliability checks and backoff branches are illustrated in the reproducible program output.

[0154] Long-term series are processed using data blocks with overlapping regions. Cross-block event merging is implemented through global index conversion, peak time tolerance, location interval overlap degree and event confidence. This can reduce the situation where events are truncated or repeatedly output near the data block boundary and support continuous processing of long-term series.

[0155] (3) Representative events and positioning results Representative anomalous spike events with different sampling frequencies and monitoring quantities were selected from existing complete localization results, and the results are shown in Table 1. The table lists the candidate width, scale range, peak time, localization interval, and two-dimensional connectivity features of each event. The above results show that the proposed method can adaptively determine the local analysis window and scale range for candidate events with different sampling frequencies and durations.

[0156] Figure 8 Taking event number 1 in Table 1 as an example, the locally centered monitoring waveform near the target event is given. Figure 8 (a) of the robust difference evidence ( Figure 8 (b) in the text, time-scale binary labeling M ( a , t ) and the two-dimensional connected region Ω associated with the candidate Figure 8 (c) and the aggregation response G ( t )( Figure 8 (d) in the text is used to describe the local multiscale analysis triggered by candidate events, and is arranged according to... M ( a , t →Ω→ G ( t The specific process for determining the peak time and positioning interval. Table 1 and Figure 8 The scale in the time scale is a discrete scale converted according to the sampling rate, and the unit is seconds.

[0157] Table 1. Representative candidate peak location results for different sampling frequencies and monitoring types.

[0158] (4) Results of measured data analysis The 45 channel-level analyses output a total of 298 deduplicated candidate events. Of these, 212 events form two-dimensional connected regions Ω associated with the candidate events, and responses are aggregated across scales. G ( t The peak time was determined; for the remaining 86 events, no effective candidate associated two-dimensional connected regions were formed, and the location was completed by using the maximum scale response according to the preset backoff rules.

[0159] In 45 channel-level analyses, the coverage ratio of unique sampling points in the local multi-scale wavelet analysis window was 1.34% on average, 0.80% on median, 0.34% on minimum, and 3.16% on maximum. This coverage ratio is the ratio of the number of unique sampling points actually entering the local multi-scale analysis to the total number of sampling points in the corresponding one-hour sequence. Therefore, within the analysis scope of this embodiment, approximately 98.66% of the sampling points do not require local multi-scale wavelet analysis on average.

[0160] The number of candidate events mentioned above is the algorithm output of this method within the given range of measured data. Since the measured data used does not provide independently annotated peak ground truth values, this embodiment is mainly used to illustrate the operability and local computational characteristics of the candidate pre-screening, local multi-scale analysis, two-dimensional connectivity localization, and backtracking process in the measured data.

[0161] As a preferred implementation, overlapping blocks of long-time sequences and deduplication of cross-block events, such as Figure 9 As shown, the process includes the following: 1. Verification data and block parameters This embodiment illustrates the process of overlapping block processing, global index conversion, and cross-block event deduplication for long-time sequences.

[0162] A long-term series with a sampling rate of 100 Hz and a total of 12001 sampling points was constructed, corresponding to a time range of 0–120.00 s. The ground truth values ​​of three single-point spikes were set at 29.98 s, 60.00 s, and 90.02 s, respectively. The spike at 29.98 s is located in the overlapping region of adjacent data blocks and is used to verify cross-block deduplication after the same anomalous event is detected repeatedly in adjacent data blocks. The spikes at 60.00 s and 90.02 s are used to verify that non-repeating events can be correctly retained during the global aggregation process.

[0163] 2. Cross-block processing steps Based on block length L B and overlap length L O Overlapping blocks are used to divide the long-term series, with a step size between adjacent data blocks. L B - L O Overlap length L O The maximum local analysis window length is preferred. L W,max Half of the maximum candidate width The larger of the two values, or a value not less than the larger value, is used to reduce the probability of candidate events located near the data block boundary being missed or incompletely located due to truncation of local waveforms. LO , L W,max , Use a consistent unit of sampling points or time.

[0164] When the length of the remaining data at the end of the sequence is less than a complete data block, the remaining data can be processed as the tail data block, or it can be padded according to a preset rule; this embodiment adopts the method of directly retaining the remaining data to form the tail data block.

[0165] Each data block independently performs candidate event generation, local multi-scale analysis, and cross-scale aggregation localization. After completing the localization within the block, the local time index of each event is converted into the global index in the original long-term series.

[0166] For events originating from adjacent data blocks, cross-block duplicate event determination conditions can be set based on peak time difference, intersection relationship of positioning intervals, or degree of overlap of positioning intervals. Peak time tolerance. δ p It can be set to a fixed number of sampling points, a fixed time, or a preset ratio of candidate width. The criteria for determining repeated events may include that the absolute value of the difference between peak times is not greater than... δ p The positioning intervals intersect and the degree of overlap of the intervals meet at least one of the preset conditions, and two or more of these conditions can be combined according to the actual application.

[0167] The degree of overlap between positioning intervals can be measured using the intersection-union ratio. R I Characterization. For two positioning intervals I 1 and I 2. The intersection and union sizes can be calculated based on the actual data representation, using the interval length or the number of sampling points contained in the interval, and the ratio of the intersection size to the union size can be used as the intersection-union ratio. R I .when R I Not lower than the preset overlap threshold ρ I When the overlap between the two positioning intervals reaches the preset condition, it is determined that the degree of overlap is equal.

[0168] For multiple cross-block event records identified as the same anomalous event, the peak time and event attributes corresponding to the event with higher confidence are retained first, and the envelope of each location interval is taken as the merged location interval. When the overall confidence of multiple events is the same, individual indicators constituting the event confidence can be further compared, such as time aggregation response. G ( tThe information includes peak magnitude or multi-sensor consistency. If the distinction cannot be made based on the above information, the retained objects can be determined according to the preset order of the global data block index. The multi-sensor consistency index is only used in the determination when multi-sensor consistency analysis is enabled. The event confidence level can be based on cross-scale over-threshold response intensity and time-aggregated response. G ( t The peak value, scale coverage, and multi-sensor consistency are determined. Alternatively, the peak time and other quantifiable continuous attributes can be weighted and fused based on event credibility; for the same batch of data, a consistent cross-block merging rule is used.

[0169] 3. Verification Results Specifically, a complete data block length of 3000 sampling points (30 s) is used, with an overlap length of 200 sampling points (2 s). The stride between adjacent data blocks is 2800 sampling points (28 s), forming a total of 5 data blocks. The first four complete data blocks cover 0–29.99 s, 28.00–57.99 s, 56.00–85.99 s, and 84.00–113.99 s, respectively, while the fifth, tail data block covers 112.00–120.00 s. Figure 9 As shown in (a) of the diagram.

[0170] After each block was independently detected and the global index was calculated, a total of 4 event records were obtained before cross-block aggregation. Among them, the same spike event at 29.98 s was detected in both data blocks because it was located in the overlapping area of ​​data block 1 and data block 2; the spikes at 60.00 s and 90.02 s each formed an event record, as shown below. Figure 9 As shown in (b) of the diagram.

[0171] Specifically, the intra-block positioning interval index in data block 1 is 2993 to 3000, and the peak index is 2999. Since the starting global index of data block 1 is 1, the converted global positioning interval is 29.92 s to 29.99 s, and the peak time is 29.98 s. The intra-block positioning interval index in data block 2 is 193 to 205, and the peak index is 199. Since the starting global index of data block 2 is 2801, the converted global positioning interval is 29.92 s to 30.04 s, and the peak time is also 29.98 s. The peak time difference between the two records is 0 s, which is not greater than the 0.10 s preset tolerance used in this embodiment, and the two positioning intervals intersect; the intersection-union ratio of the positioning intervals calculated based on discrete sampling points is... R IThe value is 8 / 13, or approximately 0.61538. Therefore, in this embodiment, based on the criterion of "intersecting positioning intervals or peak time difference not exceeding a preset tolerance," the two records are determined to be the same abnormal event. The event confidence level of both records is 0.95. The peak time and event attributes of the data block 1 record are retained according to the preset data block processing order, and the minimum envelope of the two positioning intervals, from 29.92 s to 30.04 s, is taken as the merged positioning interval. At the same time, the event source is recorded as data block 1 and data block 2.

[0172] After the above merging, three abnormal events were finally obtained.

[0173] The three output peak values ​​coincide with the true values ​​at 29.98 s, 60.00 s, and 90.02 s, respectively, with an absolute error of 0 s for all three values; the intersection-union ratio (IUU) of the corresponding positioning intervals is... R I The values ​​are 0.07692, 0.07143, and 0.07143, respectively, with average boundary errors of 0.060 s, 0.065 s, and 0.065 s, respectively. For the single-point peak true value in this embodiment, the average boundary error is defined as the average of the absolute time deviations between the left and right boundaries of the output positioning interval and the true value time. The positioning interval intersection-union ratio in this embodiment... R I The intersection and union ratios are calculated based on the discrete sampling point set. Specifically, the intersection size is the number of sampling points shared by the output positioning interval and the true value of the single-point spike, while the union size is the total number of distinct sampling points shared by both. Since the true value of a single-point spike corresponds to one sampling point, the intersection-union ratios for the three positioning intervals are 1 / 13, 1 / 14, and 1 / 14, respectively, which are 0.07692, 0.07143, and 0.07143.

[0174] 4. Results Analysis The above results demonstrate that for anomalous events located within overlapping regions of adjacent data blocks, overlapping block segmentation can reduce the risk of missed detection or incomplete localization of events near block boundaries due to local data truncation. Furthermore, it ensures that events near the boundaries in this embodiment enter the subsequent analysis process within at least one data block with relatively complete local data. Through the conversion from local index to global index and the determination and merging of cross-block events, the duplicate output of the same anomalous event can be avoided. Simultaneously, events that do not undergo cross-block duplicate detection are retained normally, indicating that the cross-block event deduplication process does not result in the erroneous deletion of non-duplicate anomalous spike events.

[0175] In this embodiment, the intersection-union ratio of the positioning interval is... R IThe lower values ​​are mainly due to the fact that the set anomaly truth value is a single sampling point, while the location interval output by this method retains the complete local response envelope around the anomaly peak for subsequent analysis or manual verification. Therefore, this embodiment is mainly used to verify overlapping block segmentation under long-term series conditions, peak location of events near block boundaries, global index conversion, and cross-block repeated event merging logic, rather than using the intersection-union ratio of location intervals under single-point truth value conditions. R I As the main performance evaluation indicator.

[0176] As a preferred implementation, a comparison of localization performance and computational efficiency with known peak truth values ​​includes the following: a. Construction of reproducible data To quantitatively evaluate peak error, positioning range, and computational efficiency, this embodiment uses random seed 20260707 to generate reproducible synthetic data. The sampling rate for each scene is 100 Hz, the time axis is 0~60.00 s, with a total of 6001 sampling points and a sampling interval of 0.01 s.

[0177] The fundamental signal consists of Gaussian noise superimposed on low-frequency components of 0.4 Hz and 1.1 Hz, with amplitudes of 0.25 and 0.12 respectively and initial phases of 0 for both components. Therefore, the fundamental deterministic signal is:

[0178] Zero-mean Gaussian noise was regenerated for each scenario, and anomalies were injected into the base signal according to the parameters listed in Table 2. The injected amplitudes in the table represent the increments to the original signal values ​​at the corresponding sampling points. z e Indicates the fast difference candidate threshold. z w This shows the preset lower limit of the wavelet discrimination threshold. The six spikes in SYN-03 are sparsely distributed in time and are not six adjacent consecutive sampling points.

[0179] In SYN-07, all three methods use channel 1 as the target channel for positioning error evaluation; channels 2-5 are only used for multi-channel consistency calculation near the candidate. The program detected two neighboring channel hits near the candidate in channel 1, with a consistency ratio of 2 / 4, or 0.5. Seven scenarios with spikes are used to calculate positioning error, while scenarios without spikes are used for false alarm detection.

[0180] Table 2 Actual generation parameters of the composite scene

[0181] b. Comparison methods and evaluation indicators Three methods were compared on the same data. Method A is a multi-scale wavelet baseline method with a preset scale for the entire signal segment. It generates finite scales on the entire signal segment according to preset minimum and maximum durations, estimates noise and thresholds scale by scale, and merges the time-scale discrimination results. Method A sets the minimum and maximum durations to 0.01 s and 1.50 s, respectively, with a nominal number of scales of 16. After logarithmic sampling, rounding of sampling points, and removal of duplicate scales, 14 scales were actually used: 0.01, 0.02, 0.03, 0.04, 0.06, 0.07, 0.10, 0.13, 0.18, 0.24, 0.32, 0.42, 0.56, and 0.75 s. At least three scales must reach the threshold at each moment, and discrimination components with a time interval not exceeding 0.05 s are merged. The minimum wavelet thresholds for each scenario are specified in Table 2. z w At the same time, the tail digits are adjusted according to the empirical tail digits of the noise response at each scale.

[0182] Method B employs fast differential candidate triggering, fixed width, fixed wavelet thresholds within a fixed threshold scene, and one-dimensional maximum aggregation. Its fixed candidate width is 0.08 s, corresponding to 8 sampling points; this width forms 9 actual scales ranging from 0.01 to 0.12 s. Each scale uses the preset wavelet thresholds listed in Table 2 of the current scene, without adaptive adjustment based on scale. During localization, the maximum value of the response at each scale is first taken at each time step, and then the peak value and continuous over-threshold intervals are determined based on the one-dimensional maximum response.

[0183] Method C employs fast differential candidate triggering, at least two classes of reliable wide evidence, backoff due to insufficient evidence, and adaptive thresholding. M ( a , t )-Ω- G ( t Two-dimensional connectivity aggregation. Method C calculates four types of width evidence: paired differential edge span, half-width continuous support width, 10%–90% cumulative energy width, and significant differential span. After eliminating widths that deviate too much from the median of evidence, the candidate width is determined by the median when there are at least two types of valid width evidence; if there is insufficient evidence, it falls back to the 0.08s preset width. The scale center is approximately half of the candidate width, and the upper and lower scale limits are approximately 1 / 3 and 3 times the scale center, respectively. Each candidate uses a maximum of 16 scales.

[0184] Method C uses the values ​​in Table 2 at each scale. z wThe threshold is set as the lower bound and adaptively adjusted using the 99.9% empirical quantile of the noise samples after removing the highest 5% of responses. For time-scale grid points exceeding the threshold, two-dimensional connected regions are formed according to 8-neighborhood relationships, prioritizing regions covering the candidate peak sampling locations; when multiple such regions exist, the region with the larger sum of out-of-threshold intensities is selected. Aggregated response. G ( t The value is obtained by equally weighted summation of the overthreshold intensities at each scale within the selected region. The final positioning interval is the minimum envelope of the time projection of the two-dimensional connected region and the fast difference candidate interval. If no effective associated region is obtained, the system regresses to the one-dimensional positioning with the largest scale.

[0185] In this set of outputs, the events that matched the 7 truth events all obtained 3 or 4 types of valid width evidence and all formed valid two-dimensional connected regions. Therefore, these matching events did not actually trigger insufficient width backoff or two-dimensional empty region backoff branches.

[0186] To verify the fallback branch when width evidence is insufficient, all candidate events generated by the SYN-05 run were further examined. One degenerate candidate event had a location interval of 33.48 s to 34.42 s, with a peak time of 34.00 s. This candidate formed two candidate edge groups, but after width reliability checks and consistency screening, only the event span determined by paired candidate edges was retained as one type of width evidence, resulting in one reliable width evidence, less than the preset two types. Therefore, the program set `width_fallback_used` to true and adopted a preset width of 0.08 s; the actual wavelet scale range determined by the preset width was 0.01 s to 0.12 s. This candidate then formed an effective two-dimensional connected region and completed the location. This degenerate candidate is not among the seven matching events used in this embodiment to calculate the true location error; it is only used to illustrate that the fallback branch when width evidence is insufficient can be executed and generate traceable width, scale, and fallback marker records.

[0187] Methods B and C share the following candidate triggering parameters: the differential statistical block length is 60 s, the merging interval of candidate edges with the same number is 0.05 s, the relative mismatch tolerance of paired edge jump amplitude is 0.85, and a maximum of 12 candidates with strong differential evidence are retained for each scenario. Local multi-scale analysis is performed on the retained candidates. The local analysis window is determined based on the candidate span and candidate width, with a basic extension length of 1.0 s on both sides of the candidate and a maximum local window length of 12 s.

[0188] If the peak distance between adjacent events after fine localization does not exceed the smaller of the estimated widths of the two events, or if the overlap length of the localization interval is not less than 50% of the shorter interval, then the events are treated as duplicate events, and events with strong evidence from comprehensive differential, wavelet, and two-dimensional regional analysis are retained.

[0189] Method A is constructed with reference to the publicly available whole-segment multi-scale wavelet processing framework and is only used as the whole-segment analysis baseline in this embodiment. It does not represent a line-by-line reproduction of all Bayesian parameterization details of a specific document.

[0190] The evaluation metrics are peak absolute error and intersection-over-union ratio of positioning intervals. R I The mean absolute error of the start and end boundaries and the coverage ratio of multi-scale wavelet deduplication analysis are used as auxiliary evaluation indicators for computational efficiency. The total running time for each method to complete the detection and localization of a single scene is recorded as an auxiliary evaluation indicator for computational efficiency.

[0191] For each scenario containing a true value, the event with the smallest absolute distance between the output peak and the true peak is selected as the matching event in principle; when the span of the true value interval is greater than 0.20 s and there is an output peak that falls into the true value interval, the output event with the smallest peak error is selected from the output events that fall into the true value interval.

[0192] This embodiment calculates the intersection-union ratio of the positioning interval according to the aforementioned discrete sampling point interval range. R I Among them, the true value interval of the isolated single-point spike of SYN-01 includes one sampling point; the true value interval of the Gaussian short pulse of SYN-06 uses the finite injection interval of 48.88~49.12 s specified by the generation program. Peak error, localization interval intersection-over-exchange ratio R I Both the boundary error and the arithmetic mean were taken over 7 scenarios containing true values.

[0193] The coverage ratio of multi-scale wavelet analysis is calculated based on the records recorded during program execution. For candidate-triggered methods, if candidate windows overlap, the cumulative processing ratio may be counted repeatedly; the deduplication analysis coverage ratio is calculated by dividing the number of unique sampling points covered by all local multi-scale wavelet windows by the total number of sampling points of the entire signal. For the entire baseline, this ratio is always 100%.

[0194] The average deduplication analysis coverage ratio in the table of this embodiment is calculated based on 8 scenarios: 7 scenarios with spikes and 1 scenario without a spike background. This metric is used to characterize the temporal sampling range entering multi-scale wavelet analysis, reflecting the effect of the candidate triggering mechanism on reducing the main multi-scale calculation range from the perspective of algorithm structure.

[0195] The actual runtime is the time elapsed from the start of the program invoking the corresponding detection method to the method returning the detection and localization results, including steps such as candidate generation, multi-scale analysis, cross-scale aggregation, and enabled consistency calculation, but excluding the time for synthetic data generation, result plotting, and file writing. Before timing, each method is warmed up once per scenario; during the formal timing phase, each scenario and each method is run 50 times, with the scenario and method order randomly shuffled in each round. The median of the 50 runs for each scenario is first taken, and then the median of the medians obtained from the eight scenarios is taken as the typical median runtime for that method. This metric is used to characterize the typical runtime level in this set of scenarios and does not indicate that every scenario achieves the same time consumption.

[0196] c. Quantitative results All three methods output at least one event in the seven scenarios with spikes, and select the corresponding event according to the above matching rules to calculate the positioning error; no event is output in scenarios without spikes.

[0197] It should be noted that the output events in the above 7 scenarios are not used as the correct detection rate or recall rate. Although the matching interval selected by Method A in SYN-04 partially overlaps with the truth interval, it does not cover the specified truth peak of 39.00 s.

[0198] Table 3. Comparison of localization performance, multi-scale analysis coverage, and typical median runtime of the three methods.

[0199] The mean peak absolute errors of methods A, B, and C are 0.07143 s, 0.08857 s, and 0.04857 s, respectively; method C reduces the error by 32.0% compared to method A and by 45.2% compared to method B. The mean positioning interval intersection ratio... R I The values ​​are 0.36, 0.41, and 0.33, respectively, and the average boundary absolute errors are 0.11 s, 0.09 s, and 0.09 s, respectively.

[0200] The average deduplication coverage rates were 100.00%, 13.20%, and 13.24%, respectively; where 13.20% and 13.24% are the average values ​​for eight scenarios, including those without sharp backgrounds. Under this average caliber, approximately 86.76% of the time sampling points in method C do not require multi-scale wavelet analysis.

[0201] As a supplement, if the deduplication analysis coverage ratio is calculated only for the 7 scenarios containing spikes, the average values ​​of Method B and Method C are 15.09% and 15.13%, respectively. Approximately 84.87% of the time sampling points in Method C do not need to be included in the multi-scale wavelet analysis. To reflect the actual operating range along with the false alarm detection scenarios in the table, this embodiment retains the average value of 8 scenarios as the table results.

[0202] In this fair repeated timing test, the typical median runtimes for methods A, B, and C were 0.038 s, 0.023 s, and 0.024 s, respectively. Method C reduces the runtime by approximately 36.6% compared to the full baseline method A; however, compared to the simpler method B, the typical median runtime of method C increases by approximately 5.0%. Therefore, method C does not aim to achieve the lowest runtime among the three methods, but rather to achieve a lower peak localization error while being significantly faster than the full baseline method A.

[0203] d. Results Analysis The results show that in the reproducible cases, the candidate triggering method reduced the coverage of multi-scale wavelet deduplication analysis from 100% of the entire baseline to approximately 13.2%; this percentage is the average result for 8 scenarios, including those without spike backgrounds. In other words, approximately 86.8% of the time sampling points of the entire sequence do not need to be included in multi-scale wavelet analysis, indicating that the candidate triggering mechanism can significantly reduce the time range covered by multi-scale computation.

[0204] The results of this fair repeated timing study also show that the typical median running time of Method C is 0.024 s, which is lower than that of Method A (0.038 s), representing a reduction of approximately 36.6%. Therefore, under the conditions of this scenario and the current program execution, Method C demonstrates a practical efficiency advantage over Method A, which covers the entire baseline, while narrowing the scope of multi-scale analysis.

[0205] Looking at the median runtime across scenarios, Method C runs slower than Method A in five scenarios: SYN-01, SYN-02, SYN-04, SYN-07, and the scenario SYN-08 without a spike background. However, it runs faster than Method A in SYN-03, SYN-05, and SYN-06. This is because SYN-03, SYN-05, and SYN-06 generate a relatively large number of candidates; SYN-05 and SYN-06 retain 9 and 12 candidates respectively, requiring repeated multi-scale analysis, adaptive processing, and two-dimensional connectivity aggregation across multiple local windows. The resulting overhead offsets the speed advantage of localized analysis over whole-process analysis. Therefore, the above typical median runtime is used to characterize the typical performance level in this set of scenarios and does not indicate that Method C runs slower than Method A in every scenario.

[0206] Method B, employing fixed width, fixed threshold, and one-dimensional maximum aggregation, has a typical median runtime of 0.023 s, approximately 5.0% lower than Method C. Method C adds reliable width estimation, adaptive thresholding, and two-dimensional connected aggregation processing, thus not requiring a runtime further lower than Method B. These results indicate that multi-scale analysis coverage primarily reflects the temporal sampling range entering the multi-scale analysis. The complete runtime is also affected by factors such as the number of candidates, the number of local windows, the number of scales, the aggregation method, and the program implementation overhead; therefore, runtime cannot be estimated solely based on coverage proportionally.

[0207] Method C has the lowest peak positioning error, but method B has the lowest positioning interval intersection-over-union ratio. R I The highest value and slightly lower boundary error. Therefore, this embodiment shows that different aggregation methods differ on different positioning metrics. The technical effect of method C is mainly reflected in: narrowing the multi-scale analysis range by using candidate triggers, and forming interrelated cross-scale anomaly evidence by utilizing time-scale two-dimensional connectivity, thereby balancing peak positioning performance and computational efficiency in this set of scenarios.

[0208] Considering that actual running time is also affected by factors such as hardware configuration, programming environment and machine load, the above typical median running time is used to illustrate the relative results under the conditions of this scenario and this program running, and is not a conclusion that method C has an absolute speed advantage under any hardware, any data and any program implementation.

[0209] The above combination Figure 1 , Figure 2 The method for identifying and locating spike anomalies in time-series signals for structural health monitoring provided in this embodiment of the invention has been described in detail. Next, the system for identifying and locating spike anomalies in time-series signals for structural health monitoring provided in this embodiment of the invention will be described in conjunction with the accompanying drawings.

[0210] Figure 10 This is a schematic diagram of the structure of a structural health monitoring time-series signal spike anomaly identification and location system according to an embodiment of the present invention, with reference to... Figure 10 The system described in this invention includes: The signal acquisition and candidate event generation module is configured to: acquire time series signals from at least one structural monitoring sensor channel, calculate the difference between adjacent sampling points, construct a differential discriminant using the robust centrality and robust differential scale of the local statistical window, extract candidate edges of abnormal peaks, associate candidate edges of abnormal peaks in the abnormal entry and exit process, and extract candidate peak events. The candidate width estimation module is configured to: select at least two width information that have passed the reliability check from the event span determined by the candidate edges of the paired abnormal peaks, the local waveform half-peak width, the local energy accumulation span, and the span between the significant difference positions on both sides of the local peak, and perform robust fusion to obtain the candidate width, thereby determining the local analysis window and the wavelet scale set; The cross-scale aggregation localization module is configured to: respond to candidate spike events, obtain multi-scale wavelet coefficients within the local analysis window corresponding to the candidate spike event according to the wavelet scale set, estimate the noise level of the multi-scale wavelet coefficients on a scale-by-scale basis and determine an adaptive threshold, acquire abnormal responses exceeding the corresponding adaptive threshold, and construct binary labels and overthreshold intensities; perform connectivity analysis on the time-scale grid points that meet the binary label conditions according to the time and scale directions to obtain a two-dimensional cross-scale response region, and aggregate the overthreshold intensities within the response region to obtain a time aggregation response, thereby determining the peak time of the abnormal spike and thus determining the localization interval.

[0211] According to embodiments of the present invention, the structural health monitoring time-series signal spike anomaly identification and localization system can correspond to the execution of the method described in the embodiments of the present invention, and the above and other operations and / or functions of each module of the structural health monitoring time-series signal spike anomaly identification and localization system are respectively for implementing Figure 1 For the sake of brevity, the corresponding processes of each method in the code will not be elaborated here.

[0212] See Figure 11 The diagram shows the structure of a computer device, which includes a processor, a communication interface, and a computer-readable storage medium. The processor, communication interface, and computer-readable storage medium are connected via a bus or other means. The communication interface is used to receive and send data. The computer-readable storage medium can be stored in the computer device's memory. The computer-readable storage medium stores computer programs, including program instructions, and the processor executes the program instructions stored in the computer-readable storage medium. The processor (or CPU, Central Processing Unit) is the computing and control core of the computer device, adapted to implement one or more instructions, specifically adapted to load and execute one or more instructions to achieve the corresponding steps in the embodiment of the structural health monitoring timing signal spike anomaly identification and location method.

[0213] This embodiment provides a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the processing system of the computer device.

[0214] Furthermore, this storage space also contains one or more instructions suitable for loading and execution by the processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device; optionally, it can also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0215] In one embodiment, the computer-readable storage medium stores one or more instructions; the processor loads and executes one or more instructions stored in the computer-readable storage medium to implement the corresponding steps in the above-described embodiment of the method for identifying and locating anomalies in timing signals during structural health monitoring.

[0216] This embodiment provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. The processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding steps in the above embodiment of the structural health monitoring timing signal spike anomaly identification and localization method.

[0217] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying and locating anomalies in time-series signals during structural health monitoring, characterized in that, include: Acquire time-series signals from at least one structural monitoring sensor channel, calculate the difference between adjacent sampling points, construct a differential discriminant using the robust centrality and robust differential scale of the local statistical window, extract candidate edges of abnormal spikes, associate candidate edges of abnormal spikes in the abnormal entry and exit processes, and extract candidate spike events. From the event span determined by the candidate edges of paired abnormal peaks, the local waveform half-peak width, the local energy accumulation span, and the span between the significant difference positions on both sides of the local peak, at least two width information that have passed the reliability check are selected and robustly fused to obtain the candidate width, thereby determining the local analysis window and the wavelet scale set. In response to candidate spike events, for the local analysis window corresponding to the candidate spike event, the multi-scale wavelet coefficients within the local analysis window are obtained according to the wavelet scale set. The noise level of the multi-scale wavelet coefficients is estimated scale by scale and an adaptive threshold is determined. Abnormal responses exceeding the corresponding adaptive threshold are obtained, and binary labels and overthreshold intensity are constructed. By performing connectivity analysis on time-scale grid points that meet the binary labeling conditions in both time and scale directions, a two-dimensional cross-scale response region is obtained. The overthreshold intensity within the response region is then aggregated to obtain the time-aggregated response, which determines the peak time of the abnormal spike and thus the location interval.

2. The method for identifying and locating structural health monitoring time-series signal spike anomalies according to claim 1, characterized in that, The method for extracting candidate edges of abnormal spikes to correlate candidate edges of abnormal entry and exit processes, and extracting candidate spike events, includes: The sampling positions where the differential discrimination value exceeds the candidate edge discrimination threshold are identified as abnormal spike candidate edges; The polarity of the candidate edges of abnormal spikes is determined based on the difference between adjacent sampling points and the robust center quantity. Abnormal spike candidate edges with a time interval not greater than a preset merging interval and the same polarity are merged into an abnormal spike candidate edge group; the abnormal spike candidate edge groups representing the abnormal entry process and the abnormal exit process have opposite polarities, and the time interval between them is between a preset minimum event span and a preset maximum event span; and according to at least one of amplitude matching relationship and local baseline change, the abnormal spike candidate edge groups representing the abnormal entry process and the abnormal exit process are associated to form candidate spike events; The amplitude matching relationship is as follows: determine the absolute amplitude between the candidate edge group of abnormal peaks representing the abnormal entry process and the candidate edge group of abnormal peaks representing the abnormal exit process. If the absolute amplitude meets the set conditions, the amplitude is considered to be matched. The local baseline change is defined as follows: when the normalized baseline change does not exceed a preset threshold, it is determined that the local baseline has been restored after the anomaly occurred.

3. The method for identifying and locating anomalies in time-series signals during structural health monitoring according to claim 1, characterized in that, The local waveform half-peak width is: a continuous half-amplitude support interval including the position of the local absolute amplitude peak; the local energy cumulative span is: the interval in which the cumulative square deviation relative to the local robust baseline reaches 10% to 90% of the total energy; the span between the significant difference positions on both sides of the local peak is: the interval on both sides of the local absolute amplitude peak that satisfies the significant difference condition. The reliability check includes: determining whether the corresponding width information is a finite positive value, whether it is within a preset allowable range, and whether the corresponding half-width support, energy accumulation interval, or significant difference position completely falls within the initial width estimation range; discarding width information that fails the reliability check; The step of selecting at least two width information values ​​that have passed reliability checks and performing robust fusion to obtain candidate widths includes: setting a width consistency tolerance coefficient. γ >1. When there are at least three types of width information involved in the consistency screening, let m W =median W j Retaining 1 / γ ≤ W j / m W ≤ γ The width information; when there are exactly two types of width information involved in the consistency screening, if max( W 1, W 2) / min( W 1, W 2)≤ γ If two width information values ​​are retained, then two width information values ​​are retained; otherwise, it is determined to be a width evidence conflict, the preset width is adopted, and a width back-off flag is output. If at least two width information values ​​are retained after consistency screening, median fusion is performed; if fewer than two width information values ​​are retained, the preset width is adopted, and a width back-off flag is output. m W For those participating in the screening W j the median of W 1 and W 2 represents the two widths when there are only two reliable widths of evidence.

4. The method for identifying and locating anomalies in time-series signals during structural health monitoring according to claim 1, characterized in that, The local analysis window includes a window center, a local analysis window length, and a local analysis window width. The window center is the time center of the candidate peak event, the middle position of the candidate edge group of abnormal peaks representing the abnormal entry and exit process, or the extreme position of the local waveform corresponding to the candidate peak event. The local analysis window length is constrained by the candidate width, the minimum window length, and the maximum window length. The local analysis window width is determined by the window center and the local analysis window length.

5. The method for identifying and locating anomalies in time-series signals during structural health monitoring according to claim 1, characterized in that, The determination of the adaptive threshold includes: At each scale, the median of the signed wavelet coefficients within the entire local analysis window is used as the robust central benchmark for noise scale estimation, and the noise scale is estimated by the absolute deviation of the wavelet coefficients from the median of the robust central benchmark. Based on the noise scale, the scale-by-scale normalized amplitude response is determined. From the scale-by-scale normalized amplitude response samples used for threshold estimation, the samples with the highest preset proportion of values ​​are excluded, and the adaptive threshold is determined based on the empirical tail quantile or the false alarm probability of the remaining samples.

6. The method for identifying and locating anomalies in time-series signals during structural health monitoring according to claim 5, characterized in that, The method involves acquiring abnormal responses exceeding a corresponding adaptive threshold and constructing binary labels and overthreshold intensities; performing connectivity analysis on time-scale grid points satisfying the binary labeling conditions along the time and scale directions to obtain a two-dimensional cross-scale response region; and aggregating the overthreshold intensities within the response region to obtain a time-aggregated response to determine the peak time of the abnormal spike, thereby determining the location interval; the method includes: Binary markers and overthreshold intensities are constructed based on the scale-normalized amplitude response and adaptive threshold. Within the local analysis window corresponding to the current candidate peak event, markers with a value of 1 are aggregated to form one or more two-dimensional connected regions according to the 8-neighborhood relationship of temporal and scale adjacency. The 8-neighborhood relationship means that the absolute value of the scale index difference between any two markers with a value of 1 is not greater than 1, the absolute value of the temporal index difference is not greater than 1, and the two index differences are not both 0. make I k For the first k The projection of a two-dimensional connected region onto the time axis I e The time interval of the candidate abnormal spike event; will satisfy I k ∩ I e ≠ The two-dimensional connected region is determined as the two-dimensional cross-scale response region. When there are multiple two-dimensional cross-scale response regions, the two-dimensional cross-scale response region associated with the current candidate peak event is determined based on at least one of the following: the degree of overlap between the time projection of each two-dimensional cross-scale response region and the time interval of the candidate peak event, the sum of the overthreshold intensity of the region, and the time distance between the time center of the region and the center of the candidate peak event. The overthreshold intensity of each scale in the two-dimensional cross-scale response region is weighted and summed to obtain the time aggregation response, and the time corresponding to the maximum value of the time aggregation response is taken as the peak time. The projection of the two-dimensional cross-scale response region on the time axis is determined as the first positioning interval, and the first positioning interval is combined with the second positioning interval determined by the candidate entry edge and exit edge by taking the envelope or fusing according to the confidence level to obtain the positioning interval of the candidate peak event.

7. A structural health monitoring time-series signal spike anomaly identification and location system, characterized in that, include: The signal acquisition and candidate event generation module is configured to: acquire time series signals from at least one structural monitoring sensor channel, calculate the difference between adjacent sampling points, construct a differential discriminant using the robust centrality and robust differential scale of the local statistical window, extract candidate edges of abnormal peaks, associate candidate edges of abnormal peaks in the abnormal entry and exit process, and extract candidate peak events. The candidate width estimation module is configured to: select at least two width information that have passed the reliability check from the event span determined by the candidate edges of the paired abnormal peaks, the local waveform half-peak width, the local energy accumulation span, and the span between the significant difference positions on both sides of the local peak, and perform robust fusion to obtain the candidate width, thereby determining the local analysis window and the wavelet scale set; The cross-scale aggregation localization module is configured to: respond to candidate spike events, obtain multi-scale wavelet coefficients within the local analysis window according to the wavelet scale set for the local analysis window corresponding to the candidate spike event, estimate the noise level of the multi-scale wavelet coefficients on a scale-by-scale basis and determine the adaptive threshold, obtain abnormal responses exceeding the corresponding adaptive threshold, and construct a binary label and over-threshold intensity. By performing connectivity analysis on time-scale grid points that meet the binary labeling conditions in both time and scale directions, a two-dimensional cross-scale response region is obtained. The overthreshold intensity within the response region is then aggregated to obtain the time-aggregated response, which determines the peak time of the abnormal spike and thus the location interval.

8. A computer device, characterized in that, A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the steps of the structural health monitoring timing signal spike anomaly identification and localization method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and execute the steps of the structural health monitoring timing signal spike anomaly identification and localization method as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps in the structural health monitoring timing signal spike anomaly identification and location method as described in any one of claims 1-6.

Images