Method for diagnosing abnormality of measuring point based on feature dimension reduction and outlier discrimination

Abstract

The application relates to a measuring point abnormality diagnosis method based on feature dimension reduction and outlier discrimination, and comprises the following steps: obtaining historical monitoring data of a fiber grating sensor; extracting a sensor feature parameter based on the historical monitoring data; performing dimension reduction processing on the sensor feature parameter; constructing a state distribution two-dimensional model of the sensor based on the dimension-reduced parameter; and performing measuring point abnormality diagnosis through an N half-neighbor value calculation method based on the two-dimensional model. The application utilizes the sensor data law of a structure health monitoring system itself, applies an algorithm to automatically identify the abnormal state of a sensor, effectively indicates an abnormal sensor measuring point, and saves the maintenance workload of the sensor.

Description

Technical Field

[0001] This invention relates to the field of fiber optic grating sensor anomaly diagnosis technology, and in particular to a measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier discrimination. Background Technology

[0002] In structural health monitoring systems, a large number of sensors need to be deployed to collect structural status information. Due to the large number of sensors and the complex working environment, they may malfunction due to external damage, natural aging, environmental changes, etc. during long-term use, making the collected data unreliable and thus affecting the performance of the monitoring system. Therefore, the maintenance and abnormal condition diagnosis of the sensors themselves are also very important.

[0003] This invention primarily addresses anomaly diagnosis using fiber optic grating sensors. Fiber optic grating sensors are a type of fiber optic sensor. The sensing process based on fiber optic gratings acquires sensing information by modulating the wavelength of the fiber Bragg grating using external physical parameters; it is a wavelength-modulated fiber optic sensor. In large-scale space structure health monitoring systems, fiber optic grating strain sensors are frequently used to monitor structural stress, enabling real-time assessment of the structural stress state, and have numerous application examples.

[0004] Traditional fault diagnosis for such sensors often relies on threshold judgment of the sensor's output signal. Alternatively, it involves manual observation of the historical output data trajectory of each sensor to determine its condition, supplemented by on-site inspections. In contrast, pattern recognition-based methods typically require system modeling of the sensor, simulation of various faults to obtain fault data, and subsequent algorithm training, necessitating substantial prior knowledge.

[0005] Existing structural health monitoring systems utilize the inherent regularity of fiber Bragg grating sensors to diagnose sensor faults. However, they have strict requirements for feature selection, and the fusion index is one-dimensional, making it impossible to distinguish between different types of faults. Furthermore, the visualization of sensor state distribution is poor.

[0006] Large-scale space structure health monitoring systems deploy numerous fiber optic grating sensors with high sampling frequencies and massive amounts of data. However, their operation is susceptible to external interference, resulting in abnormal value jumps. A single threshold jump judgment is insufficient to accurately determine whether the sensor's operating status is normal. On the other hand, while manual observation yields relatively accurate results, the workload is excessive for a large number of measurement points, making timely and effective completion difficult. Pattern recognition-based methods require extensive simulation experiments, training samples, and prior knowledge. The method proposed in patent CN113029242A requires strict selection of signal features. If multiple features representing a certain fault are repeatedly selected, collinearity of feature vectors can lead to inaccurate identification of other faults, affecting the accuracy of the identification results. Furthermore, this method uses a one-dimensional fusion index, resulting in a limited indication effect, inability to distinguish between different fault types, and poor visualization of the judgment results.

[0007] Traditional unsupervised discrimination methods based on data density, even those suitable for this example, all have significant drawbacks. For instance, the K-means algorithm requires pre-specifying the number of clusters; the DBSCAN method, a noisy density-based clustering method, requires pre-setting neighborhood thresholds and the number of core points, making it difficult to identify faulty samples when the proportion changes; the Isolation Forest method also requires pre-setting the proportion of faulty points, making it impractical in real-world applications with unknown anomalies; the LOF anomaly detection algorithm is also heavily reliant on the number of neighboring points and anomaly thresholds, leading to errors when multiple faulty points of the same type are clustered together; and other commonly used algorithms largely depend on parameter tuning. In this example, the number of sensor fault types and the proportion of faulty sensors are unpredictable, making it impossible for traditional methods to achieve universal adaptive discrimination in applications. Summary of the Invention

[0008] The purpose of this invention is to propose a measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier discrimination to solve the problems existing in the prior art.

[0009] To achieve the above objectives, the present invention provides the following solution:

[0010] Acquire historical monitoring data from fiber Bragg grating sensors;

[0011] Based on the historical monitoring data, sensor characteristic parameters are extracted;

[0012] The sensor's characteristic parameters are subjected to dimensionality reduction processing;

[0013] Based on the reduced-dimensional parameters, a two-dimensional model of the sensor's state distribution is constructed.

[0014] Based on the two-dimensional state distribution model, anomaly diagnosis of measuring points is performed by calculating the N semi-nearest neighbor value.

[0015] Optionally, acquiring historical monitoring data from the fiber Bragg grating sensor includes:

[0016] Determine the time period range for the historical monitoring data that needs to be acquired;

[0017] Based on the time period range, the database is read cyclically by referring to the sensor number list, the data is segmented, and the historical monitoring data of a single sensor within the time period range is extracted.

[0018] Useless attributes are removed from historical monitoring data, and only the wavelength values ​​corresponding to the sensor strain values, as well as the measurement point number and time label, are retained;

[0019] In the retained data, the data corresponding to the measurement points where the channel is closed are removed.

[0020] Optionally, the sensor characteristic parameters include: sample length, data standard deviation, data variance, time-domain energy value of the measurement point signal, mean-removed waveform range, mean-removed waveform peak value, mean-removed peak value index, mean-removed kurtosis index, original signal waveform index, original signal pulse factor, original signal margin coefficient, signal main frequency component peak value, and signal main frequency component frequency.

[0021] Optionally, before performing dimensionality reduction processing on the sensor feature parameters, the method further includes:

[0022] The sensor characteristic parameters are normalized.

[0023] Optionally, normalizing the sensor characteristic parameters includes:

[0024] The feature values ​​of each sensor measurement point are organized into a matrix form to obtain an m×n dimensional feature value matrix, where m is the number of sensor measurement points and n is the number of extracted feature values. Each row in the feature value matrix represents all feature values ​​of a sensor measurement point, and each column represents the distribution of a certain feature value among all measurement points.

[0025] Normalize all feature data by taking each column of the matrix as a set of data, i.e., each feature value.

[0026] Optionally, dimensionality reduction of the normalized parameters includes:

[0027] The normalized parameters are then centered.

[0028] Calculate the covariance matrix based on the centralized data;

[0029] Perform eigenvalue decomposition on the covariance matrix to calculate eigenvalues ​​and eigenvectors;

[0030] Take the eigenvectors corresponding to the first few eigenvalues ​​as principal components and construct a projection matrix, where each column is an eigenvector;

[0031] The centered data is projected into a subspace composed of the first few feature vectors to obtain the dimensionality-reduced dataset.

[0032] Optionally, based on a two-dimensional model, anomaly diagnosis of measurement points is performed using the N-half-nearest neighbor calculation method, including:

[0033] Calculate the N-half neighbor value for each measuring point and obtain a list of N-half neighbor values ​​for all measuring point samples;

[0034] Sort the N semi-nearest neighbor list in ascending order, take the first preset number of values ​​in the list, and obtain a new list;

[0035] Calculate a preset threshold based on the new list;

[0036] The N semi-nearest neighbor value of each measuring point is compared with a preset threshold. If the N semi-nearest neighbor value of a measuring point is greater than the preset threshold, the measuring point is identified as a faulty measuring point.

[0037] Optionally, calculating the N semi-nearest neighbor value for each measuring point includes:

[0038] For each measuring point, calculate the Euclidean distance from that measuring point to all other measuring points, sort these distances in ascending order, and take the N / 2th distance value as the N semi-nearest neighbor value of that measuring point; where N is the total number of measuring points, and N / 2 is rounded down.

[0039] Optionally, the method for calculating the preset threshold based on the new list includes:

[0040]

[0041] Where, k T Here, T is the threshold discrimination coefficient, M is the number of samples in the new list, i is the index of the new list, and d is the threshold discrimination coefficient. si The N semi-nearest neighbor value is the i-th measurement point in the new list.

[0042] The beneficial effects of this invention are as follows:

[0043] This invention utilizes the inherent patterns in sensor data from the structural health monitoring system and applies algorithms to automatically identify abnormal sensor states, effectively indicating abnormal sensor measurement points and saving on sensor maintenance workload.

[0044] By extracting relevant features that can characterize various sensor faults and reducing the dimensionality of multiple feature parameters, the impact of feature vector collinearity is reduced, resulting in more accurate identification results.

[0045] By using a two-dimensional plane to plot the feature distribution of sensor measurement points after dimensionality reduction, the visualization of sensor status is improved, allowing for intuitive observation of the clustering and outlier states of different sensor measurement point features. Furthermore, sensor measurement points with the same type of fault tend to cluster in the two-dimensional plane of the identification results, which helps in fault type classification and targeted inspection.

[0046] By utilizing the different data distribution densities, the dimensionality-reduced sensor data is scanned and calculated to automatically identify whether the overall data is clustered and to automatically identify outlier measurement points, thus achieving unsupervised automatic judgment and diagnosing sensor anomalies. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram of the measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier discrimination according to an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the original sample of the measurement point in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the frequency domain characteristics of the measurement points after FFT transformation according to an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram showing the distribution of all sensor feature values ​​in an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram showing the sensor state distribution at different measuring points according to an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of the diagnostic results according to an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] Fiber Bragg grating strain gauges output wavelength values ​​as their raw signal, which have a linear correlation with the measured strain value. Due to the inherent temperature sensitivity of fiber Bragg grating sensors, the wavelength values ​​output by normal strain monitoring sensors show a significant correlation with ambient temperature and exhibit periodic variations. Furthermore, within the same spatial structure, structural stress changes also show a correlation due to the structural integrity. Therefore, for fiber Bragg grating sensor measurement points deployed on the same monitoring object, their output data exhibit statistically similar patterns. Faulty points are often a minority, and their data often represent outliers in various statistical indicators.

[0057] Based on this characteristic, this invention analyzes and extracts all feature values ​​that characterize the sensor state and are independent of the sensor's initial value. These features are then subjected to dimensionality reduction processing, and the sensor distribution is visualized using a two-dimensional plane. Furthermore, an algorithm is applied to automatically identify outlier sensor measurement points based on their density clustering relationship, thus diagnosing abnormal states. Sensors with the same fault type also tend to cluster together in the planar distribution, which aids in the classification of sensor fault types.

[0058] like Figure 1 As shown, the measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier discrimination proposed in this embodiment includes:

[0059] Acquire historical monitoring data from fiber Bragg grating sensors;

[0060] Based on historical monitoring data, extract sensor characteristic parameters;

[0061] The sensor's characteristic parameters are subjected to dimensionality reduction processing;

[0062] Based on the reduced-dimensional parameters, a two-dimensional model of the sensor's state distribution is constructed.

[0063] Based on the two-dimensional state distribution model, anomaly diagnosis of measuring points is performed by calculating the N semi-nearest neighbor value.

[0064] Furthermore, acquiring historical monitoring data from the fiber Bragg grating sensor includes:

[0065] Determine the time period range for the historical monitoring data that needs to be acquired;

[0066] Based on the time period range, the database is read cyclically by referring to the sensor number list, the data is segmented, and historical monitoring data of a single sensor within the time period range is extracted.

[0067] Useless attributes are removed from historical monitoring data, and only the wavelength values ​​corresponding to the sensor strain values, as well as the measurement point number and time label, are retained;

[0068] In the retained data, the data corresponding to the measurement points where the channel is closed are removed.

[0069] Specifically, in this embodiment, the extraction of historical monitoring data includes the following:

[0070] In structural health monitoring systems, the system records and stores data in a database, resulting in a massive amount of data. To accurately analyze the sensor status, it is necessary to extract a period of historical data over a longer monitoring period.

[0071] First, determine the time period for the data used in this analysis. For example, if you want to extract one month of historical monitoring data, ensure that the historical period does not involve significant structural changes, the structural stress signal is stable, and the main variable load is temperature load. Then, refer to the sensor number list, cyclically read the database, segment the data, and extract the historical data for each sensor within that time period. If the sensor data is not found in the list, it means that the measuring point channel is closed. Record and output the sensor number whose channel is closed for verification with the monitoring software. Remove useless attributes by discarding data from the sensor records that is irrelevant to fault diagnosis, retaining only the wavelength values ​​corresponding to the sensor strain values, the measuring point number, and the time label. Save the data locally. Based on the retrieved data, remove the measuring points with closed channels, renumber the existing data sequentially, and retain the new numbers and the corresponding sensor measuring points. Store each measuring point as a separate data file as the original sample.

[0072] Example of original sample at measurement point: Figure 2 As shown.

[0073] Furthermore, the sensor characteristic parameters include: sample length, data standard deviation, data variance, time-domain energy value of the measured point signal, mean-removed waveform range, mean-removed waveform peak value, mean-removed peak value index, mean-removed kurtosis index, original signal waveform index, original signal pulse factor, original signal margin coefficient, peak value of the main frequency component of the signal, and frequency of the main frequency component of the signal.

[0074] Specifically, in this embodiment, all sensor feature parameters are extracted iteratively. Theoretically, all features that are only related to the shape of the signal waveform and are independent of the initial value of the signal can be extracted as recognition features. The main features and extraction formulas are as follows:

[0075] (1) Sample length:

[0076] In most health monitoring systems, fiber Bragg grating sensors often use threshold detection for data recording and storage. Abnormal values ​​exceeding the threshold are not recorded in the database, or due to channel errors, software errors, or other reasons, the database fails to record every sampling information in detail. In this example, data is extracted by time period. For different sensors, the final sample length extracted from the database is highly likely to be inconsistent (almost inevitable). Therefore, the final sample length extracted for each measurement point is one of the important criteria for judging whether the sensor is functioning properly. The sample length for a single measurement point is denoted as:

[0077] L = N = length(x).

[0078] (2) Standard deviation of data:

[0079] Standard deviation reflects the dispersion of data values ​​at measurement points and can be used as a basis for assessing signal stability. The data file saved in the previous step is read sequentially and cyclically. For a single measurement point, the input data is the sensor output and the strain signal wavelength x. i =x1,x2,x3,...,x N ;

[0080] In the formula, i is the data point number, i = 1, 2, 3, ..., N. N is the sample length.

[0081] The formula for calculating the standard deviation at each measurement point is:

[0082] (3) Data variance:

[0083]

[0084] In the formula For the input signal x i Sequence mean.

[0085] (4) Time-domain energy value of the measurement point signal:

[0086] The time-domain energy value can reflect the intensity of signal output fluctuations and is a good indicator for faulty sensors with no significant output change. The data is first de-meaned before the energy value at that measurement point is calculated.

[0087] The formula for calculating the time-domain energy value is:

[0088]

[0089] (5) Mean-removed waveform range:

[0090]

[0091] (6) Mean-removed waveform peak value: represents the maximum fluctuation of the waveform;

[0092]

[0093] (7) Mean-to-peak indicator:

[0094]

[0095] (8) Mean-removed kurtosis index:

[0096]

[0097] (9) Original signal waveform specifications:

[0098]

[0099] (10) Original signal pulse factor:

[0100]

[0101] (11) Original signal margin coefficient:

[0102]

[0103] (12) Peak value of the main frequency component of the signal:

[0104] The principal component of the frequency domain of the measured signal reflects the period of the main load at the measured point. A normally functioning fiber optic grating sensor should reflect the ambient temperature load, which typically varies over a period of approximately one day. Performing a Fourier transform on the sample signal yields the signal's spectral function.

[0105]

[0106] In the formula, N is the original sample length.

[0107] In computer processing, to reduce computational load, Fast Fourier Transform (FFT) is used here to obtain the signal spectrum and normalize it.

[0108] After obtaining the signal spectrum, iterate through F(k) (k=0,1,2,…,N-1) to find the maximum value max(F(k)), which is the principal frequency component, and obtain its corresponding frequency kmax.

[0109] (13) Frequency of the main frequency component of the signal:

[0110] Example of frequency domain characteristics after FFT transformation of the measurement point is shown in the figure below. Figure 3 As shown.

[0111] In the previous step, kmax is the frequency corresponding to the main frequency component of the signal.

[0112] Furthermore, in this embodiment, the sensor characteristic parameters are first normalized;

[0113] Then, the normalized parameters are subjected to data dimensionality reduction processing.

[0114] Furthermore, normalizing the sensor characteristic parameters includes:

[0115] The feature values ​​of each sensor measurement point are organized into a matrix form to obtain an m×n dimensional feature value matrix, where m is the number of sensor measurement points and n is the number of extracted feature values. Each row in the feature value matrix represents all feature values ​​of a sensor measurement point, and each column represents the distribution of a certain feature value among all measurement points.

[0116] Normalize all feature data by taking each column of the matrix as a set of data, i.e., each feature value.

[0117] Furthermore, dimensionality reduction of the normalized parameters includes:

[0118] The normalized parameters are then centered.

[0119] Calculate the covariance matrix based on the centralized data;

[0120] Perform eigenvalue decomposition on the covariance matrix to calculate eigenvalues ​​and eigenvectors;

[0121] Take the eigenvectors corresponding to the first few eigenvalues ​​as principal components and construct a projection matrix, where each column is an eigenvector;

[0122] The centered data is projected into a subspace composed of the first few feature vectors to obtain the dimensionality-reduced dataset.

[0123] Specifically, in this embodiment, the normalization of sensor feature data includes the following:

[0124] First, each feature value is numerically normalized to avoid excessive weight differences between feature values ​​due to differences in quantity.

[0125] Suppose there are m sensor measurement points, and each sensor measurement point extracts n feature values, requiring m > n, then the sensor's m×n dimensional feature value matrix can be obtained:

[0126]

[0127] In each row [p i1 ,p i2 ,...,p in ] represents all feature values ​​of the i-th sensor measurement point.

[0128] Each column represents the distribution of a certain eigenvalue across all measurement points. The vector of the j-th eigenvalue is denoted as: v j =[p 1j ,p 2j ,...,p mj ]

[0129] Normalize all feature data, treating each column (i.e., each feature value) as a group:

[0130]

[0131] After normalization, the new sensor feature matrix is ​​obtained:

[0132]

[0133] The distribution of all sensor eigenvalues ​​is as follows: Figure 4 As shown.

[0134] Specifically, in this embodiment, feature data dimensionality reduction includes the following:

[0135] Due to the large number of features, it is difficult to quickly identify the sensor status. Therefore, dimensionality reduction is performed on the feature values ​​extracted from all sensor measurement points.

[0136] Principal component analysis (PCA) is applied to reduce the dimensionality of the feature data to two dimensions.

[0137] The PCA calculation steps are as follows:

[0138] (1) Centralize the dataset:

[0139]

[0140] in, Let represent the mean of the i-th eigenvalue.

[0141] (2) Calculate the covariance matrix:

[0142]

[0143] Where, σ ij Let represent the covariance between the i-th eigenvalue and the j-th eigenvalue.

[0144] (3) Calculate eigenvalues ​​and eigenvectors:

[0145] Perform eigenvalue decomposition on the covariance matrix H to find its eigenvalues ​​and eigenvectors. The eigenvectors satisfy the following:

[0146] Hv=λv

[0147] Where λ is the eigenvalue corresponding to the eigenvector v, we get the eigenvalues ​​λ1, λ2, ..., λ nand the corresponding feature vectors v1, v2, ..., v n .

[0148] (4) Calculate the principal components:

[0149] For the eigenvector v and the corresponding eigenvalue λ, their PCA components are calculated as follows:

[0150] Take the eigenvectors corresponding to the first two eigenvalues ​​as principal components and construct a projection matrix W, where each column is an eigenvector.

[0151] Projecting the dataset onto the subspace formed by the first two feature vectors yields the dimensionality-reduced PCA principal components:

[0152] PCA = X0W

[0153] The resulting PCA is the dimensionality-reduced dataset, consisting of two columns and m rows, with each row representing the two dimensionality-reduced feature values ​​of a measurement point.

[0154] Specifically, in this embodiment, a two-dimensional model of the sensor state distribution is constructed based on the preprocessed parameters, i.e., a two-dimensional plane of the sensor state distribution is drawn; including the following:

[0155] Based on the eigenvalues ​​in the dimensionality-reduced data, with the first column as the x-axis and the second column as the y-axis, a sensor state distribution map for different measuring points is plotted; that is, with the first column of PCA eigenvalues ​​as the x-axis and the second column as the y-axis, the sensor state distribution map for different measuring points is plotted as follows. Figure 5 As shown:

[0156] According to the state distribution Figure 5 This allows for easy observation of whether any abnormal measuring points exist among all measuring points. It is evident that when no faulty measuring points are present, the sensor status distribution is relatively uniform and dispersed.

[0157] When there are faulty measuring points, because a certain characteristic value of the faulty measuring point is significantly different from that of the normal measuring points, it becomes more prominent in the aforementioned processing flow. In the two-dimensional plane distribution, the normal measuring points will be clustered together, while the faulty measuring points will be clearly outliers, and the states of the same type of faulty measuring points will be close to each other in the two-dimensional plane.

[0158] In this embodiment, after achieving two-dimensional visualization, the state distribution of the sensor can be observed manually. However, it is not possible to use a computer to automatically determine which measurement points are abnormal. The last step of this embodiment (through the calculation method of N semi-nearest neighbor values) is to achieve manual judgment through automatic calculation, which can adaptively complete the abnormal diagnosis of outlier measurement points.

[0159] Furthermore, based on the two-dimensional model, the anomaly diagnosis of measurement points is performed using the N-half-nearest neighbor calculation method, including:

[0160] Calculate the N-half neighbor value for each measuring point and obtain a list of N-half neighbor values ​​for all measuring point samples;

[0161] Sort the list of N semi-nearest values ​​in ascending order, take the first preset number of values ​​in the list, and obtain a new list;

[0162] Calculate the preset threshold based on the new list;

[0163] The N semi-nearest neighbor value of each measuring point is compared with a preset threshold. If the N semi-nearest neighbor value of a measuring point is greater than the preset threshold, the measuring point is identified as a faulty measuring point.

[0164] Specifically, in this embodiment, the automatic fault detection method based on a two-dimensional plane includes the following:

[0165] The premise and basic idea of ​​this embodiment is that, in practical applications, faulty sensors are necessarily a minority, meaning that more than half of the sensors are functioning correctly. When faulty sensors exist, through the aforementioned dimensionality reduction steps, the distribution of normal sensors will inevitably cluster together, thereby enabling unsupervised diagnosis. This aligns with the fundamental laws of natural aging and random damage of sensors in sensor systems, meaning that only a few sensors will inevitably fail prematurely.

[0166] (1) First, to identify its distribution status, the concept of "N-half nearest neighbor distance" (NHNND) is defined. This is the distance from a given measuring point to the nearest (N / 2)th point when the total number of measuring points is N. Here, N / 2 should be rounded down. The distance calculation uses Euclidean distance, which is a two-dimensional space in this method. The formula for the distance between two points is:

[0167]

[0168] Where (x1, y1) are the two feature parameters of the first point, and (x2, y2) are the feature parameters of the other point.

[0169] The process of calculating the N semi-nearest neighbor value of a point is as follows: First, calculate the distance from the point to all other measurement points, including the point itself, and store the N distance values ​​as a distance list for the point. Then, sort the values ​​in the distance list in ascending order, i.e., from smallest to largest. Take the value at index N / 2 (rounded down) in the sorted distance list, which is the N semi-nearest neighbor value of the point.

[0170] This can be understood as follows: in a two-dimensional plane, when a circle is drawn with the measuring point as the center and N semi-nearest values ​​as the radius, at least N / 2 measuring points will be included in the drawn circle or fall on the edge of the circle.

[0171] (2) Next, by calculating the N-half neighbor value for each measuring point, a list of N-half neighbor values ​​for all measuring point samples can be formed: D = d1, d2, ... d N In the formula d i That is, the N semi-nearest neighbor value of the i-th measurement point.

[0172] (3) Define the normal threshold calculation method in this embodiment. Sort the list of all N semi-nearest neighbor values ​​of all measurement points in ascending order, and take the first M = N / 2 (rounded down) values ​​after sorting the list to obtain the list:

[0173] D s =d s1 ,d s2 ,...d sM M = int(N / 2)

[0174] The formula for calculating the threshold is:

[0175]

[0176] In the formula, k T k is the threshold discrimination coefficient. In this embodiment, k T Taking 3 shows good adaptability to samples of different quantities and fault types.

[0177] (4) Fault diagnosis, defining fault diagnosis methods:

[0178]

[0179] In the formula, S i Let d represent the state determination result of the i-th sensor measurement point. i This represents the N semi-nearest neighbor value of the i-th sensor.

[0180] (5) When the method is applied to sensor fault diagnosis, the identification results are highly consistent with the actual state for different proportions and types of faulty sensors in the measurement point sample. The faulty sensor can be accurately identified under different circumstances.

[0181] The diagnostic results are plotted in two dimensions. A red circle is drawn on the graph, scaled proportionally to the N-half nearest neighbor value of the faulty measuring point, and the sensor number of the faulty measuring point is marked. This clearly shows the fault location and its relative severity. The diagnostic results are as follows: Figure 6 As shown.

[0182] This embodiment utilizes the inherent sensor data patterns of the structural health monitoring system to extract historical data, automatically extract sensor signal characteristics, automatically identify sensor state distribution, effectively indicate abnormal sensors, and save on manual inspection workload.

[0183] It extracts comprehensive features, effectively detects more types of sensor faults, and reduces the dimensionality of multiple feature parameters, thereby reducing the impact of feature vector collinearity and making the recognition results more accurate.

[0184] The distribution of sensor measurement point features after dimensionality reduction is plotted using a two-dimensional plane, which has excellent visualization effect and allows for intuitive observation of the clustering and outlier states of different sensor measurement point features.

[0185] Using the invented method, the sensor data after dimensionality reduction can be calculated to automatically identify whether there are outliers in the measurement point status in an unsupervised state, and outliers can be identified as fault states.

[0186] The algorithm is highly adaptable and can effectively identify different proportions of sensor fault points. In practical applications, the algorithm parameters do not require much manual intervention.

[0187] Sensor measurement points with the same type of fault cluster together in the two-dimensional plane of the identification results, which helps to classify the fault type and conduct targeted inspections.

[0188] Based on the magnitude of the N semi-nearest neighbor value, the relative severity of the fault at the sensor measurement point can be observed intuitively.

[0189] Compared to outlier detection algorithms such as LOF and deep learning, this algorithm is simpler in calculation and requires less computation while ensuring accurate results.

[0190] This embodiment is based on the fact that the raw data output by the fiber optic strain sensor in the health monitoring system has a uniform pattern, and the abnormal sensor signal is significantly different from the normal sensor in certain characteristics. Based on the historical output of the sensor, an automatic diagnosis method for abnormal sensors is established.

[0191] The proposed diagnostic method first standardizes all extracted fault feature values ​​to eliminate unit and order-of-magnitude limitations between different feature values. Then, it applies a data dimensionality reduction method to transform them into two-dimensional data and performs sensor status diagnosis based on the distribution pattern of data points in the two-dimensional plane.

[0192] A sensor state visualization method is proposed, which uses dimensionality-reduced two-dimensional features as the ordinates and abscissas of a two-dimensional plane to visualize the state distribution of a large number of sensors.

[0193] Based on the principle that faulty measuring points are always few, a new method for detecting outliers, the N-semi-nearest neighbor method, is proposed to calculate the distribution of measuring point states.

[0194] This embodiment proposes an innovative N-semi-nearest neighbor diagnostic method.

[0195] The main objectives of this embodiment are:

[0196] By utilizing the patterns in sensor data from the structural health monitoring system itself, algorithms are applied to automatically identify abnormal sensor conditions, effectively indicating abnormal sensor measurement points and saving sensor maintenance workload.

[0197] By extracting relevant features that can characterize various sensor faults and reducing the dimensionality of multiple feature parameters, the impact of feature vector collinearity is reduced, resulting in more accurate identification results.

[0198] By using a two-dimensional plane to plot the feature distribution of sensor measurement points after dimensionality reduction, the visualization of sensor status is improved, allowing for intuitive observation of the clustering and outlier states of different sensor measurement point features. Furthermore, sensor measurement points with the same type of fault tend to cluster in the two-dimensional plane of the identification results, which helps in fault type classification and targeted inspection.

[0199] By utilizing the different data distribution densities, the dimensionality-reduced sensor data is scanned and calculated to automatically identify whether the overall data is clustered and to automatically identify outlier measurement points, thus achieving unsupervised automatic judgment and diagnosing sensor anomalies.

[0200] The algorithm's adaptability is improved, and it can effectively identify different proportions of sensor fault points. In practical applications, the algorithm parameters do not require much manual intervention.

[0201] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier detection, characterized in that, include: Acquire historical monitoring data from fiber Bragg grating sensors; Based on the historical monitoring data, sensor characteristic parameters are extracted; The sensor feature parameters are subjected to dimensionality reduction processing; Based on the reduced-dimensional parameters, a two-dimensional model of the sensor's state distribution is constructed. Based on the two-dimensional state distribution model, anomaly diagnosis of measuring points is performed by calculating the N semi-nearest neighbor value. Based on a two-dimensional model, anomaly diagnosis of measurement points is performed using the N-half-nearest neighbor (NNN) calculation method, including: Calculate the N-half neighbor value for each measuring point and obtain a list of N-half neighbor values ​​for all measuring point samples; Sort the N semi-nearest neighbor list in ascending order, take the first preset number of values ​​in the list, and obtain a new list; Calculate a preset threshold based on the new list; The N-half neighbor value of each measuring point is compared with a preset threshold. If the N-half neighbor value of a measuring point is greater than the preset threshold, the measuring point is identified as a faulty measuring point. Calculating the N semi-nearest neighbor values ​​for each measuring point includes: For each measuring point, calculate the Euclidean distance from that measuring point to all other measuring points, sort these distances in ascending order, and take the N / 2th distance value as the N half-neighbor value of that measuring point; where N is the total number of measuring points, and N / 2 is rounded down; The method for calculating the preset threshold based on the new list includes: in, Here, T is the threshold discrimination coefficient, M is the preset threshold, and i is the index of the new list. The N semi-nearest neighbor value of the i-th measurement point in the new list.

2. The measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier discrimination according to claim 1, characterized in that, Historical monitoring data obtained from fiber Bragg grating sensors includes: Determine the time period range for the historical monitoring data that needs to be acquired; Based on the time period range, the database is read cyclically by referring to the sensor number list, the data is segmented, and the historical monitoring data of a single sensor within the time period range is extracted. Useless attributes are removed from historical monitoring data, and only the wavelength values ​​corresponding to the sensor strain values, as well as the measurement point number and time label, are retained; In the retained data, the data corresponding to the measurement points where the channel is closed are removed.

3. The measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier discrimination according to claim 1, characterized in that, The sensor characteristic parameters include: sample length, data standard deviation, data variance, time-domain energy value of the measured signal, mean-removed waveform range, mean-removed waveform peak value, mean-removed peak value index, mean-removed kurtosis index, original signal waveform index, original signal pulse factor, original signal margin coefficient, peak value of the main frequency component of the signal, and frequency of the main frequency component of the signal.

4. The measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier discrimination according to claim 1, characterized in that, Before performing dimensionality reduction processing on the sensor feature parameters, the following steps are also included: The sensor characteristic parameters are normalized.

5. The measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier discrimination according to claim 4, characterized in that, Normalizing the sensor characteristic parameters includes: The feature values ​​of each sensor measurement point are organized into a matrix form to obtain an m×n dimensional feature value matrix, where m is the number of sensor measurement points and n is the number of extracted feature values. Each row in the feature value matrix represents all feature values ​​of a sensor measurement point, and each column represents the distribution of a certain feature value among all measurement points. Normalize all feature data by taking each column of the matrix as a set of data, i.e., each feature value.

6. The measurement point anomaly diagnosis method based on feature dimensionality reduction and outlier discrimination according to claim 4, characterized in that, Dimensionality reduction of normalized parameters includes: The normalized parameters are then centered. Calculate the covariance matrix based on the centralized data; Perform eigenvalue decomposition on the covariance matrix to calculate eigenvalues ​​and eigenvectors; Take the eigenvectors corresponding to the first few eigenvalues ​​as principal components and construct a projection matrix, where each column is an eigenvector; The centered data is projected into a subspace composed of the first few feature vectors to obtain the dimensionality-reduced dataset.

Images