Railway perimeter intrusion recognition method based on vibration wave spectrum analysis enhanced in frequency domain

Abstract

The present invention relates to the technical field of railway perimeter intrusion recognition, and discloses a method for railway perimeter intrusion recognition by vibration spectrum analysis with frequency domain enhancement, including: S01, constructing an original data set, performing spectrum analysis on the sample signals in the original data set to obtain a spectrum feature data set; S02, training a classifier using the spectrum feature data set; S03, performing spectrum analysis on the actual vibration signal to obtain actual vibration spectrum features; S04, using the trained classifier to recognize the actual vibration spectrum features; the spectrum analysis includes: S100, performing windowing and framing processing on the amplitude sequence of the signal according to the sampling rate of the vibration optical fiber sensor, etc.; this method realizes real-time monitoring of foreign object intrusion in the railway perimeter, and can also monitor foreign object intrusion in bad weather with low visibility such as strong wind and heavy rain.

Description

Technical Field

[0001] The present invention relates to the technical field of railway perimeter intrusion recognition, and in particular to a method for recognizing railway perimeter intrusion by vibration spectrum analysis with frequency domain enhancement. Background Art

[0002] As of the end of 2022, the operating mileage of the national railway reached up to 1.549 million kilometers, and the pressure of railway operation safety prevention is increasing. In complex mountainous areas, there are risks of falling rocks, landslides or debris flows invading the railway operation boundary. When a road crosses over a railway line or runs parallel to a high-speed line, there is a danger that vehicles and on-vehicle items may fall and invade the boundary. There are risks of intrusion such as passengers or items on the platform falling under the platform, and the bridge pier being accidentally impacted by targets such as bridges and debris flows. In addition, operating railways are vulnerable to the influence of the external environment, and there are risks of floating objects such as greenhouse sheds and color steel panel houses invading the line. Above all, the perimeter intrusion of foreign objects poses a great risk to railway operation safety. In the lightest case, the train stops midway, and in the most serious case, derailment occurs, endangering the safety of passengers. The current railways are equipped with certain natural disaster and perimeter intrusion equipment and systems to ensure the safety of the operation environment.

[0003] Defects and deficiencies of the prior art: At present, the railway perimeter intrusion prevention facilities mainly rely on pulsed and tension electronic fences. The layout cost and maintenance cost are high, and the layout has high requirements for infrastructure such as power supply and communication. Its layout scope is mainly applied to stations and sections where roads and railways run parallel. The layout conditions are limited, and large-area layout cannot be completed. The current railway integrated video surveillance perimeter intrusion recognition equipment belongs to two independent units, and data sharing and communication cannot be carried out. The application of video mainly relies on personnel for identification, and automatic target identification cannot be achieved. For target intrusion recognition based on video, there is a problem that it cannot be normally applied under bad weather such as heavy fog and heavy rain. The distributed vibration optical fiber system can effectively make up for the video gap under bad weather. At present, the railway perimeter intrusion recognition devices and methods based on distributed vibration optical fibers are easily misreported due to the influence of railway operation and the surrounding environment. In addition, they can only locate the intrusion position and cannot identify the type of intrusion event. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the purpose of the present invention is to provide a method for recognizing railway perimeter intrusion by vibration spectrum analysis with frequency domain enhancement, which reduces the layout cost and maintenance cost and has good recognition effect.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] A method for identifying railway perimeter intrusion by vibration spectrum analysis with frequency domain enhancement, comprising: S01, constructing an original data set, performing spectrum analysis on the sample signals in the original data set to obtain a spectrum feature data set; S02, training a classifier using the spectrum feature data set; S03, performing spectrum analysis on the actual vibration signal to obtain actual vibration spectrum features; S04, identifying the actual vibration spectrum features using the trained classifier; the spectrum analysis includes: S100, performing windowing and framing processing on the amplitude sequence of the signal according to the sampling rate of the vibration optical fiber sensor; S200, calculating the band spectrum entropy and short-time kurtosis based on the windowing and framing results of the amplitude sequence, and performing two-parameter and two-threshold endpoint detection based on the calculation results to determine the occurrence location of the intrusion event; S300, calculating the variance of the change rate of the band energy probability density of the disturbance segment amplitude sequence obtained according to the endpoint detection result to determine the band where the vibration event energy is concentrated; S400, performing band enhancement on the band where the vibration event energy is concentrated according to the linear combination between the moving mean difference algorithms under different differential step sizes; S500, extracting spectrum features based on the original amplitude sequence and the result of band enhancement.

[0007] In the present invention, preferably, before S01, it further includes: S00, dividing the disturbance events into large-range events and small-range events based on the duration and influence range of the actual vibration signal, classifying the events with a duration greater than 15 seconds and a vibration range greater than 50 meters as large-range events, and otherwise classifying them as small-range events.

[0008] In the present invention, preferably, the formula for the windowing and framing processing in S100 is:

[0009] ,

[0010] where, assuming that the time-domain signal obtained by the monitoring point on the sensing optical fiber is , is the th signal of the th frame, is the signal of the th frame, is the sampling rate of the optical fiber sensor, is the length of the amplitude sequence included in each frame, is the length of the overlapping sequence between adjacent frames, is the window function.

[0011] In the present invention, preferably, the duration of each frame in S100 is , the overlapping rate between adjacent frames is , the length of the amplitude sequence included in each frame , the length of the overlapping sequence between adjacent frames ; The window function uses a Hamming window, which is:

[0012] .

[0013] In the present invention, preferably, the S200 includes:

[0014] S201, performing a discrete Fourier transform on each frame of signal to calculate the spectrum of each frame of signal. The formula is:

[0015] , ,

[0016] where, is the th signal within the th frame after windowing and frame division;

[0017] S202, equally spacing the signal into a number of frequency bands according to the sampling rate, and calculating the logarithmic energy of each frequency band of each frame of signal. The formula is:

[0018] ;

[0019] where, is a normal quantity, with a value of , is the logarithmic energy of the th frequency band of the th frame of signal, is the number of spectral lines of each frequency band;

[0020] S203, calculating the probability density and frequency band spectral entropy of the energy of each frequency band by using the logarithmic energy of each frequency band. The probability density formula is:

[0021] ,

[0022] The frequency band spectral entropy formula is:

[0023] ,

[0024] where, is the probability density of the energy of the th frequency band of the th frame of signal, is the logarithmic energy of the th frequency band of the th frame of signal, is a normal quantity, with a value of 0.5, is the frequency band spectral entropy of the th frame;

[0025] S204. Calculate the mean value of the signal amplitude, and calculate the kurtosis of the signal by using the mean value of the signal amplitude. The formula is:

[0026] ,

[0027] where, is the kurtosis of the th frame signal, is the mean value of the amplitude of the th frame signal, is the th signal of the th frame, is the length of the amplitude sequence included in each frame;

[0028] S205. Determine the low spectral entropy threshold and the high spectral entropy threshold in the frequency band spectral entropy, and determine the low kurtosis threshold in the kurtosis;

[0029] S206. Determine the points less than the low spectral entropy threshold on the frequency band spectral entropy as the perturbation points;

[0030] S207. Calculate the time interval between adjacent perturbation points;

[0031] S208. Attribute the adjacent perturbation points with a time interval less than the time interval threshold to the same perturbation segment;

[0032] S209. Continuously attribute the perturbation segments with a time interval less than the time interval threshold to the same perturbation segment until the time intervals between all perturbation segments are not less than the time interval threshold;

[0033] S210. Determine the points greater than the low kurtosis threshold on the kurtosis curve as the kurtosis boundary points, and continuously extend the perturbation segments with a time interval less than the time interval threshold from the kurtosis boundary points until the time intervals between all perturbation segments and all kurtosis boundary points are not less than the time interval threshold;

[0034] S211. Continuously attribute the perturbation segments with a time interval less than the time interval threshold to the same perturbation segment until the time intervals between all perturbation segments are not less than the time interval threshold, and merge the different perturbation segments with time overlap into the same perturbation segment;

[0035] S212. Discard the single perturbation segments with a duration less than the duration threshold.

[0036] In the present invention, preferably, the S300 includes: S301. Calculate the energy probability density of each frequency band of each frame signal of the perturbation segment; S302. Perform differential processing on the energy probability density of each frequency band; S303. Calculate the variance of the differential result of the energy probability density of each frequency band and the mean value of the variance; S304. Determine the frequency bands with the variance of the differential result of the energy probability density greater than the mean value of the variance as the frequency bands where the perturbation event energy is concentrated.

[0037] In the present invention, preferably, the S400 includes: S401, performing a moving average process on the energy concentrated frequency band; S402, setting several different differential step lengths to perform a moving difference process on the moving average result respectively; S403, setting the weight of the energy concentrated frequency band to be higher than the weights of the remaining frequency bands, and linearly combining the moving average difference results with different differential step lengths to obtain an amplitude sequence with enhanced frequency bands.

[0038] In the present invention, preferably, the S500 includes: S501, extracting the short-time amplitude variance, short-time kurtosis, variance of the energy change rate of each frequency band, and variance of the frequency band spectral entropy of the amplitude sequence of the signal; S502, extracting spectral features based on the amplitude sequence with enhanced frequency bands, and the spectral features include the number of wave peaks, the number of wave valleys, the time difference between adjacent wave peaks, and the amplitude difference between a wave peak and its adjacent wave valley within a unit time.

[0039] In the present invention, preferably, the steps for identifying wave peaks and wave valleys in S502 are as follows:

[0040] S5021, identifying through the second-order difference result of the amplitude sequence with enhanced frequency bands, identifying that the second-order difference greater than 0 is a wave valley, and identifying that the second-order difference less than 0 is a wave peak. The calculation formula for the second-order difference is:

[0041] ,

[0042] where the amplitude sequence of the signal after enhancement in the perturbation section is , is the th amplitude after enhancement of the signal in the perturbation section, is the second-order difference of;

[0043] S5022, calculating the wave peak and wave valley saliency, and removing the wave peaks with wave peak and wave valley saliency lower than the saliency threshold. The calculation formula for the wave peak and wave valley saliency is:

[0044] ,

[0045] where is the wave peak and wave valley saliency, is the wave peak amplitude, is the wave valley amplitude.

[0046] In the present invention, preferably, the classifier adopts the random forest algorithm.

[0047] This method uses processing such as endpoint detection and frequency-domain enhancement to achieve the extraction of spectral data such as railway operation and rockfall intrusion in frequency bands. The frequency-band extraction is based on spectral features such as short-time amplitude variance, short-time kurtosis, variance of energy change rate in each frequency band, and variance of spectral entropy in the frequency band, and the random forest method is used for intrusion event classification. Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. Realize real-time monitoring of foreign object intrusion on the railway perimeter, and can also monitor foreign object intrusion in bad weather with low visibility such as strong winds and heavy rains;

[0049] 2. According to the energy proportion distribution of intrusion events, the prominent and feature extraction of the frequency band with concentrated energy in the sequence are realized;

[0050] 3. Realize the identification of railway perimeter vibration events, and can identify train operation, harsh environments and various intrusion events. Description of the Drawings

[0051] Figure 1 It is a flowchart of a method for identifying railway perimeter intrusion by vibration spectrum analysis with frequency-domain enhancement according to an embodiment of the present invention.

[0052] Figure 2 It is a schematic diagram of the amplitude curve of the time-domain signal in a method for identifying railway perimeter intrusion by vibration spectrum analysis with frequency-domain enhancement according to an embodiment of the present invention.

[0053] Figure 3 It is a flowchart of spectrum analysis in a method for identifying railway perimeter intrusion by vibration spectrum analysis with frequency-domain enhancement according to an embodiment of the present invention.

[0054] Figure 4 It is a schematic diagram of windowing and framing in a method for identifying railway perimeter intrusion by vibration spectrum analysis with frequency-domain enhancement according to an embodiment of the present invention.

[0055] Figure 5 It is a flowchart of S200 in spectrum analysis in a method for identifying railway perimeter intrusion by vibration spectrum analysis with frequency-domain enhancement according to an embodiment of the present invention.

[0056] Figure 6 It is a schematic diagram of dividing the disturbance section of the curve by S208 in a method for identifying railway perimeter intrusion by vibration spectrum analysis with frequency-domain enhancement according to an embodiment of the present invention.

[0057] Figure 7 It is the highlighting and suppression effects of different N values on each frequency signal in a method for identifying railway perimeter intrusion by vibration spectrum analysis with frequency-domain enhancement according to an embodiment of the present invention.

[0058] Figure 8 It is a flowchart of S300 in spectrum analysis in a method for identifying railway perimeter intrusion by vibration spectrum analysis with frequency-domain enhancement according to an embodiment of the present invention.

[0059] Figure 9 It is a flowchart of S400 for spectral analysis in the method for identifying railway perimeter intrusion by vibration spectral analysis with frequency domain enhancement according to an embodiment of the present invention.

[0060] Figure 10 It is a schematic diagram of the curve after frequency band enhancement in the method for identifying railway perimeter intrusion by vibration spectral analysis with frequency domain enhancement according to an embodiment of the present invention.

[0061] Figure 11 It is a flowchart of S500 for spectral analysis in the method for identifying railway perimeter intrusion by vibration spectral analysis with frequency domain enhancement according to an embodiment of the present invention.

[0062] Figure 12 It is a flowchart of S502 for spectral analysis in the method for identifying railway perimeter intrusion by vibration spectral analysis with frequency domain enhancement according to an embodiment of the present invention.

[0063] Figure 13 It is a schematic diagram of the characteristic distribution of short-time peak value and short-time amplitude variance for two intrusion events in the method for identifying railway perimeter intrusion by vibration spectral analysis with frequency domain enhancement according to an embodiment of the present invention.

[0064] Figure 14 It is a flowchart of spectral analysis in the method for identifying railway perimeter intrusion by vibration spectral analysis with frequency domain enhancement according to another embodiment of the present invention.

[0065] Figure 15 It is a logical flowchart of random forest classification in the method for identifying railway perimeter intrusion by vibration spectral analysis with frequency domain enhancement according to another embodiment of the present invention. Specific Embodiments

[0066] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0067] It should be noted that when a component is referred to as being "fixed to" another component, it can be directly on the other component or there may also be an intermediate component. When a component is considered to be "connected to" another component, it can be directly connected to the other component or there may be an intermediate component at the same time. When a component is considered to be "disposed on" another component, it can be directly disposed on the other component or there may be an intermediate component at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are only for the purpose of illustration.

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this invention belongs. The terms used in the specification of this invention are for the purpose of describing specific embodiments only and are not intended to limit the invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0069] Please also refer to Figures 1 to 3 , a preferred embodiment of the present invention provides a method for identifying railway perimeter intrusion by frequency-domain enhanced vibration spectrum analysis, including:

[0070] S01, construct an original data set, perform spectrum analysis on the sample signals in the original data set, and obtain a spectrum feature data set.

[0071] The original data set consists of a certain number of sample signals, and the sample signals are amplitude time-domain signals obtained by a vibration optical fiber sensor. As Figure 2 shown, it shows the amplitude curve of the time-domain signal. The sample signals have been completed in identification and are divided into different event types. The sample signals cannot be effectively identified by the classifier directly. Therefore, it is necessary to perform spectrum analysis to form the spectrum features of the sample signals, and then be identified by the classifier, so as to achieve the purpose of training the classifier. Correspondingly, the original data set forms a spectrum feature data set composed of the spectrum features of each sample signal through spectrum analysis.

[0072] S02, use the spectrum feature data set to train the classifier.

[0073] S03, perform spectrum analysis on the actual vibration signal to obtain the actual vibration spectrum feature.

[0074] The actual vibration signal is also an amplitude time-domain signal obtained by a vibration optical fiber sensor, showing the amplitude curve of the time-domain signal. The actual vibration signal cannot be effectively identified by the classifier directly. Therefore, it is necessary to perform spectrum analysis to form the spectrum features of the actual vibration signal, and then be identified by the classifier, so as to determine what kind of event it is.

[0075] S04, use the trained classifier to identify the actual vibration spectrum feature.

[0076] The trained classifier can automatically and accurately identify the actual vibration spectrum feature, thus replacing the manual identification of vibration events based on the actual vibration signals transmitted by the vibration optical fiber sensor.

[0077] Among them, spectrum analysis is performed in both S01 and S03. As Figure 3 shown, the spectrum analysis specifically includes:

[0078] S100, window the amplitude sequence of the signal according to the sampling rate of the vibration optical fiber sensor and perform frame segmentation processing.

[0079] In this step, there are multiple choices for the specific function and method used in windowing and frame segmentation processing. The description here is only one example. To maintain the continuity of the signal and enable smooth transition between each frame of the signal, the overlapping frame segmentation method is used to process the time-domain signal. Assume that the time-domain signal obtained at a certain monitoring point on the sensing optical fiber within a certain period of time is , and the sampling rate of the optical fiber sensor is , the duration of each frame is , the overlapping rate between adjacent frames is , and the window function uses the Hamming window. Then, denote the signal of the th frame as , then there is:

[0080]

[0081] Among them, is the length of the amplitude sequence included in each frame, is the length of the overlapping sequence between adjacent frames, is the Hamming window function, then there is:

[0082]

[0083]

[0084]

[0085] The effect of windowing and frame segmentation is as shown in Figure 4 .

[0086] S200, calculate the frequency band spectral entropy and short-time kurtosis according to the results of windowing and frame segmentation of the amplitude sequence, and perform two-parameter and two-threshold endpoint detection based on the calculation results to determine the location where the intrusion event occurs.

[0087] Specifically, as shown in Figure 5 , S200 includes:

[0088] S201, perform discrete Fourier transform on each frame of the signal to calculate the spectrum of each frame of the signal. The formula is:

[0089] , ;

[0090] Among them, is the th signal within the th frame after windowing and frame segmentation. First, perform discrete Fourier transform on each frame of the voice signal to calculate its spectrum.

[0091] S202. Divide the signal into several frequency bands at equal intervals according to the sampling rate, and calculate the logarithmic energy of each frequency band of each frame of the signal.

[0092] For example, according to the sampling rate the signal is divided into 10 frequency bands at equal intervals, then each frequency band has spectral lines, so the logarithmic energy of the th frequency band of the

[0093]

[0094] frame signal is: where is a normal quantity, generally taking the value of , is the logarithmic energy of the th frequency band of the th frame signal, and

[0095] S203. Calculate the probability density and frequency band spectral entropy of the energy of each frequency band by using the logarithmic energy of each frequency band.

[0096] Correspondingly, the probability density of the energy of the th frequency band of the th frame signal is:

[0097]

[0098] where is a normal quantity, generally taking 0.5.

[0099] And the frequency band spectral entropy of the th frame is obtained:

[0100]

[0101] where is the probability density of the energy of the th frequency band of the th frame signal, is the logarithmic energy of the th frequency band of the th frame signal, is a normal quantity, taking the value of 0.5, is the th frame's frequency band spectral entropy.

[0102] S204. Calculate the mean value of the signal amplitude, and calculate the kurtosis of the signal by using the mean value of the signal amplitude.

[0103] S204. Calculate the mean value of the signal amplitude, and calculate the kurtosis of the signal by using the mean value of the signal amplitude.

[0104] If is the mean value of the amplitude of the -th frame signal, then the kurtosis of the -th frame signal is:

[0105]

[0106] where is the kurtosis of the -th frame signal, is the mean value of the amplitude of the -th frame signal, is the -th signal of the -th frame, is the length of the amplitude sequence contained in each frame.

[0107] S205. Determine the low spectral entropy threshold and the high spectral entropy threshold in the frequency band spectral entropy, and determine the low kurtosis threshold in the kurtosis.

[0108] For the endpoint detection based on the frequency band spectral entropy and the short-time kurtosis, first determine a relatively low spectral entropy low threshold and a relatively high spectral entropy high threshold in the frequency band spectral entropy , and determine a relatively low kurtosis low threshold in the kurtosis. These three thresholds are all determined after on-site experimental tests after installing the vibrating optical fiber on-site. When the system is installed, the characteristic values in the quiet state and the characteristic values of the simulation experiment are measured, and then set.

[0109]

[0109] S206. Determine the points less than the low spectral entropy threshold on the frequency band spectral entropy as the disturbance points.

[0110] Here, it is considered that the points on the curve less than are the disturbance points. Denote the disturbance points as in sequence from left to right. There are disturbance points:

[0111]

[0112] S207. Calculate the time intervals between adjacent disturbance points.

[0113] Calculate the time intervals between adjacent disturbance points, denoted as , and there is:

[0114]

[0115] S208. Group adjacent disturbance points with a time interval less than the time interval threshold into the same disturbance segment.

[0116] If the time interval between adjacent disturbance points is less than the time interval threshold (e.g., 1 - 10 s), it is considered that the two are caused by the same disturbance event, group them into the same disturbance segment, and denote it as , where is the start time and end time of the th disturbance segment to complete the first - level decision. There are disturbance segments, as shown in Figure 6 .

[0117] S209. Continuously group disturbance segments with a time interval less than the time interval threshold into the same disturbance segment until the time intervals between all disturbance segments are not less than the time interval threshold.

[0118] Continuing with as the reference, search for points on that are less than respectively from the start point and end point of the disturbance segment to both sides, and denote them as and respectively. If there exists or with a time distance from the disturbance segment less than the time interval threshold , then replace the time of the start point or end point of the disturbance segment with the moment of or and repeat the above operation until there is no with a time distance of a disturbance segment less than or to complete the second - level decision.

[0119] S210. Determine the points on the kurtosis curve that are greater than the lower kurtosis threshold as kurtosis boundary points, and continuously extend the disturbance segments with a time interval less than the time interval threshold from the kurtosis boundary points until the time intervals between all disturbance segments and all kurtosis boundary points are not less than the time interval threshold.

[0120] S210. Determine the points on the kurtosis curve that are greater than the lower kurtosis threshold as kurtosis boundary points, and continuously extend the disturbance segments with a time interval less than the time interval threshold from the kurtosis boundary points until the time intervals between all disturbance segments and all kurtosis boundary points are not less than the time interval threshold.

[0121] Similarly, using the kurtosis curve as the reference, search for points on that are greater than respectively from the start point and end point of the disturbance segment after the second - level decision to both sides, and denote them as and If there exists or the time interval to the disturbance segment is less than the time interval threshold then replace the time of the starting point or the ending segment corresponding to the disturbance segment with or the moment, and repeat the above operation until there is no time interval less than of or to complete the three - level decision, and record each obtained disturbance segment as .

[0122] S211. Continuously classify the disturbance segments with a time interval less than the time interval threshold into the same disturbance segment until the time intervals between all disturbance segments are not less than the time interval threshold, and merge different disturbance segments with time overlap into the same disturbance segment.

[0123] After completing the first - level, second - level, and third - level decisions, several disturbance segments will be obtained If the time interval between adjacent disturbance segments is less than the time interval threshold then merge the two segments into one disturbance segment. If there is time overlap between the disturbance segments of adjacent sampling points, it is considered that the two sampling points are caused by the same intrusion event, and the disturbance segments with overlap should be merged.

[0124] S212. Discard single disturbance segments with a duration less than the duration threshold.

[0125] If the duration of a single disturbance segment is less than the duration threshold (usually between 300 ms and 1000 ms), then it is considered that this disturbance is caused by noise and the disturbance segment is discarded.

[0126] S300. According to the disturbance segment amplitude sequence obtained from the endpoint detection result, calculate the variance of the change rate of the frequency - band energy probability density of the disturbance segment amplitude sequence to determine the frequency band where the vibration event energy is concentrated.

[0127] Assume that the time - domain signal of a certain sampling point is , first perform moving average processing on it to smooth its noise:

[0128]

[0129] Then perform moving difference processing on the moving average result to reflect its amplitude change:

[0130]

[0131] where is the size of the smoothing window, is the step size of the difference calculation, and for and There are:

[0132]

[0133]

[0134] Therefore, there are , and the remaining sequence lengths are both after the two are processed, and the results of the two only differ in the amplitude values, and there is no influence on the signal-to-noise ratio, frequency distribution, etc.

[0135] It can be assumed that the sampling frequency is , and there is a sinusoidal vibration with a frequency of on the sensing optical fiber. Then a complete sinusoidal vibration will be expanded on sampling signals. Then the vibration phase difference between adjacent sampling signals is . When the moving differential processing with a step size of is performed on it, the sum of the absolute values of the results within one vibration period is:

[0136]

[0137] It can be seen that when the sampling frequency and the vibration frequency are fixed, the sum of its absolute values is only related to the differential step size . Therefore, by adjusting the value, the purpose of suppressing or highlighting the signal with a frequency of can be achieved, as shown in Figure 7 . Therefore, when the sampling frequency is and , the suppression or highlighting effect of its result on the signal with a frequency of to is recorded as the row vector . The length of the row vector is , then there is:

[0138]

[0139] Therefore, by linearly combining the moving average differential results of different values, the purpose of highlighting or suppressing a certain frequency band signal can be achieved, that is, to solve:

[0140]

[0141]

[0142] Among them, is the matrix composed of the calculation results of different values, is the coefficient of the linear combination to be solved, is the weight of the frequency band to be highlighted or suppressed, ' is the transposed vector of.

[0143] Specifically, as Figure 8 shown, S300 includes:

[0144] S301, calculating the energy probability density of each frequency band of each frame signal in the disturbance segment.

[0145] According to the above analysis, first determine the energy concentration segment of the vibration event through the frequency band characteristics. First, calculate the th frame signal of the disturbance segment th energy probability density of the frequency band , and its calculation method is shown in formula (7).

[0146] S302, performing differential processing on the energy probability density of each frequency band.

[0147] Then perform differential processing on the energy probability density of each frequency band to calculate its change rate:

[0148]

[0149] S303, calculating the variance of the differential result of the energy probability density of each frequency band and the mean value of the variance.

[0150] Then calculate the variance of the differential result of the energy probability density of the th frequency band:

[0151]

[0152] where is the total number of frames of the differential sequence in the disturbance segment, is the mean value of the differential sequence of the energy probability density of the th frequency band.

[0153] Then calculate the mean value of the variances of each frequency band .

[0154]

[0155] S304, determining the frequency bands with the variance of the differential result of the energy probability density greater than the mean value of the variance as the energy concentration frequency bands of the disturbance event.

[0156] Here, it is considered that the frequency bands with the variance of the frequency band greater than the mean value are the main energy concentration frequency bands (energy concentration frequency bands) of the disturbance event, denoted as .

[0157] The S400 enhances the frequency band of the energy concentration frequency band of the vibration event through the linear combination between the moving average difference algorithms under different differential step lengths.

[0158] According to the foregoing analysis, when the sampling frequency and the vibration frequency are fixed, the sum of their absolute values is only related to the differential step length . The frequency band enhancement of the energy concentration frequency band can be carried out according to the foregoing method. Specifically, as Figure 9 shown, S400 includes:

[0159] S401, performing a moving average process on the energy concentration frequency band.

[0160] The moving average process can be carried out according to formula (12).

[0161] S402, setting several different differential step lengths to perform a moving difference process on the moving average result respectively.

[0162] The moving difference process can be carried out according to formula (13). Since the results of performing differential processes with different differential step lengths are required later, this step calculates the differential results of multiple different differential step lengths.

[0163] S403, setting the weight of the energy concentration frequency band to be higher than the weights of the remaining frequency bands, linearly combining the moving average difference results of different differential step lengths to obtain an amplitude sequence after frequency band enhancement.

[0164] Set the weight of the energy concentration frequency band to be higher than the weights of the remaining frequency bands, and highlight the main frequency bands where the disturbance events are concentrated through the linear combination between the moving difference mean results of different values. This step can be carried out according to formulas (18) and (19). The curve of frequency band enhancement is as Figure 10 shown.

[0165] S500, extracting spectral features based on the original amplitude sequence and the results of frequency band enhancement.

[0166] Specifically, as Figure 11 shown, S500 includes:

[0167] S501, extracting the short-time amplitude variance, short-time kurtosis, variance of the energy change rate of each frequency band, and variance of the frequency band spectral entropy of the amplitude sequence of the signal;

[0168] First, extract the short-time amplitude variance , short-time kurtosis of the signal based on the amplitude sequence of the original data, the variance of the energy change rate of each frequency band and the variance of the frequency band spectral entropy :

[0169]

[0170]

[0171]

[0172]

[0173] S502. Extract spectral features based on the amplitude sequence after frequency band enhancement. The spectral features include the number of wave peaks, the number of wave valleys, the time difference between adjacent wave peaks, and the amplitude difference between a wave peak and its adjacent wave valley within a unit time.

[0174] The short-time peak value and the short-time amplitude variance characteristics of two intrusion events are distributed as Figure 13 shown. Through learning, the classifier can accurately identify. Specifically, as Figure 12 shown, the steps for identifying wave peaks and wave valleys in S502 are as follows:

[0175] S5021. Identify wave peaks and wave valleys through the second-order difference result of the amplitude sequence after frequency band enhancement. Identify those with a second-order difference greater than 0 as wave valleys , and identify those with a second-order difference less than 0 as wave peaks . Assume that the amplitude sequence after signal enhancement in a certain disturbance section is , then the calculation formula for the second-order difference is:

[0176] ,

[0177] where the amplitude sequence after signal enhancement in the disturbance section is , is the th amplitude after signal enhancement in the disturbance section, is 's second-order difference.

[0178] S5022. Calculate the saliency of wave peaks and wave valleys, and remove the wave peaks with saliency lower than the saliency threshold. To reduce the influence of noise, calculate the saliency of wave peaks and wave valleys , and set a threshold to remove the influence of noise. The saliency is calculated using the height difference between a wave peak and its adjacent wave valley. The calculation formula for the saliency of wave peaks and wave valleys is:

[0179] ,

[0180] where is the saliency of wave peaks and wave valleys, is the wave peak amplitude, is the trough amplitude.

[0181] In a preferred embodiment of the present invention, as Figure 14 shown, before S01, the vibration wave spectrum analysis railway perimeter intrusion recognition method with frequency domain enhancement further includes:

[0182] S00, dividing the disturbance events into large - scale events and small - scale events based on the duration and influence range of the actual vibration signal. An event with a duration greater than 15 seconds and a vibration range greater than 50 meters is classified as a large - scale event, otherwise it is classified as a small - scale event.

[0183] Before identifying the intrusion event through spectrum analysis, the events can be classified according to the influence range and duration of the disturbance section of the vibration event represented by the actual vibration signal transmitted by the vibration optical fiber sensor. If the duration within the disturbance section of the vibration event is greater than 15 seconds and there is a moment when the vibration range exceeds 50 meters, then the disturbance event is classified as a large - scale event, otherwise it is classified as a small - scale event. For large - scale events, it is considered that the event types include three types of events such as train movement, landslides, debris flows and other large - scale geological disasters, as well as bad weather such as strong winds and heavy rains. For small - scale events, it is considered that the events include four types of events such as a small amount of stones falling onto the railway track and the perimeter area of the railway track, pedestrian or animal intrusion, pedestrians damaging the railway or the protective net, and gentle breeze weather.

[0184] In a preferred embodiment of the present invention, the classifier adopts the random forest algorithm. The process of generating the random forest model is as follows: First, randomly sample with replacement from the original dataset to form multiple sampling sample sets. Then, train the sub - decision trees using the CART algorithm. Repeat constructing multiple sub - decision trees by the same criterion for the decision trees. Finally, form a random forest. Use the random forest model to discriminate and classify new data. The final result is determined by voting, and the category with the most votes is used as the final classification result. The flow of the random forest classification algorithm is as Figure 15 shown.

[0185] The above description is a detailed description of the preferred and feasible embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. Any equivalent changes or modifications made under the technical spirit disclosed by the present invention should fall within the scope of the patent covered by the present invention.

Claims

1. A railway perimeter intrusion identification method based on frequency domain enhanced vibration spectrum analysis, characterized in that: include: S01, constructing an original data set, performing spectrum analysis on sample signals in the original data set, and obtaining a spectrum feature data set; S02, training the classifier using the spectral feature dataset; S03, performing spectrum analysis on the actual vibration signal to obtain actual vibration spectrum characteristics; S04, using the trained classifier to identify the actual vibration spectrum features; The spectral analysis includes: S100, performing windowing and framing processing on the amplitude sequence of the signal according to the sampling rate of the vibration optical fiber sensor; S200, calculating the frequency band spectrum entropy and short-time kurtosis according to the amplitude sequence windowing and framing results, and performing dual-parameter and dual-threshold endpoint detection based on the calculation results to determine the location of the intrusion event; S300, calculating the variance of the frequency band energy probability density change rate of the disturbance segment amplitude sequence according to the disturbance segment amplitude sequence obtained from the endpoint detection result, and determining the frequency band where the vibration event energy is concentrated; S400, frequency band enhancement of the frequency band where the vibration event energy is concentrated is performed based on the linear combination of the moving mean difference algorithms under different difference step sizes; S500, spectral feature extraction based on the original amplitude sequence and frequency band enhancement results; The S200 includes: S201, performing discrete Fourier transform on each frame signal to calculate the spectrum of each frame signal, the formula is: ， ， in, After windowing and framing The first A signal; S202, dividing the signal into a number of frequency bands at equal intervals according to the sampling rate, and calculating the frequency band logarithmic energy of each frequency band of each frame signal, the formula is: ； in, is a normal quantity, and its value is , For the The frame signal The logarithmic energy of the frequency band, is the number of spectral lines for each frequency band; S203, using the frequency band logarithmic energy of each frequency band, calculate the probability density of the energy of each frequency band and the frequency band spectral entropy, and the probability density formula is: ， The frequency band spectral entropy formula is: ， in, For the The frame signal The probability density of the energy in each frequency band is For the The frame signal The logarithmic energy of the frequency band, is a normal quantity, with a value of 0.

5. For the The frequency band spectral entropy of the frame; S204, calculating the mean value of the signal amplitude, and using the mean value of the signal amplitude to calculate the signal kurtosis, the formula is: ， in, For the The kurtosis of the frame signal, For the The mean value of the frame signal amplitude, For the Frame No. A signal, is the length of the amplitude sequence contained in each frame; S205, determining a spectral entropy low threshold and a spectral entropy high threshold in the frequency band spectral entropy, and determining a kurtosis low threshold in the kurtosis; S206, determining a point less than a low threshold of spectral entropy from the frequency band spectral entropy as a disturbance point; S207, calculating the time interval between adjacent disturbance points; S208, assigning adjacent disturbance points whose time interval is less than a time interval threshold to the same disturbance segment; S209, continuously classifying disturbance segments whose time intervals are less than the time interval threshold as the same disturbance segment until the time intervals between the disturbance segments are no less than the time interval threshold; S210, determining a point on the kurtosis curve that is greater than a low kurtosis threshold as a kurtosis boundary point, and continuously extending disturbance segments whose time intervals with the kurtosis boundary point are less than the time interval threshold to the corresponding kurtosis boundary point, until the time intervals between each disturbance segment and each kurtosis boundary point are not less than the time interval threshold; S211, continuously classifying disturbance segments whose time intervals are less than the time interval threshold as the same disturbance segment until the time intervals between the disturbance segments are no less than the time interval threshold, and merging different disturbance segments with overlapping time into the same disturbance segment; S212, discarding a single disturbance segment whose duration is less than a duration threshold.

2. The method for identifying railway perimeter intrusions by frequency domain enhanced vibration spectrum analysis according to claim 1 is characterized in that: Before S01, it also included: S00, based on the duration and impact range of the actual vibration signal, the disturbance events are divided into large-scale events and small-scale events. Events with a duration greater than 15 seconds and a vibration range greater than 50 meters are classified as large-scale events, otherwise they are classified as small-scale events.

3. The method for identifying railway perimeter intrusions by frequency domain enhanced vibration spectrum analysis according to claim 1 is characterized in that: The formula used for the windowing and framing process in S100 is: ， It is assumed that the time domain signal obtained by the monitoring point on the sensing optical fiber is , For the Frame No. A signal, For the Frame signal, is the sampling rate of the fiber optic sensor, is the length of the amplitude sequence contained in each frame, is the overlap length of the sequence between adjacent frames, is the window function.

4. The method for identifying railway perimeter intrusions by frequency domain enhanced vibration spectrum analysis according to claim 3 is characterized in that: The duration of each frame in S100 is The overlap rate between adjacent frames is , the amplitude sequence length contained in each frame , the overlapping sequence length between adjacent frames ; The window function adopts a Hamming window, which is: 。 5. The method for identifying railway perimeter intrusions by frequency domain enhanced vibration spectrum analysis according to claim 1 is characterized in that: The S300 includes: S301, calculating the energy probability density of each frequency band of each frame signal of the disturbance segment; S302, performing differential processing on the energy probability density of each frequency band; S303, calculating the variance and the mean of the variance of the energy probability density difference results of each frequency band; S304: Determine a frequency band in which the variance of the energy probability density difference result is greater than the mean of the variance as a frequency band in which the energy of the disturbance event is concentrated.

6. The method for identifying railway perimeter intrusions by frequency domain enhanced vibration spectrum analysis according to claim 1 is characterized in that: The S400 includes: S401, performing moving average processing on the energy concentrated frequency band; S402, setting a number of different difference step sizes to perform moving difference processing on the moving average results respectively; S403, setting the weight of the energy concentrated frequency band to be higher than the weights of the remaining frequency bands, linearly combining the moving mean difference results with different difference step sizes, and obtaining an amplitude sequence after the frequency band is enhanced.

7. The method for identifying railway perimeter intrusions by frequency domain enhanced vibration spectrum analysis according to claim 1 is characterized in that: The S500 includes: S501, extracting the short-time amplitude variance, short-time kurtosis, energy change rate variance of each frequency band, and frequency band spectral entropy variance of the amplitude sequence of the signal; S502, extracting spectrum features based on the frequency band enhanced amplitude sequence, wherein the spectrum features include the number of peaks and troughs per unit time, the time difference between adjacent peaks, and the amplitude difference between a peak and its adjacent trough.

8. The method for identifying railway perimeter intrusions by frequency domain enhanced vibration spectrum analysis according to claim 7 is characterized in that: The steps for identifying peaks and troughs in S502 are: S5021, through the second-order difference result of the amplitude sequence after frequency band enhancement, it is determined that the second-order difference is greater than 0 as a trough, and the second-order difference is less than 0 as a peak. The calculation formula of the second-order difference is: ， Among them, the amplitude sequence of the enhanced disturbance signal is: , is the first signal after the disturbance segment is enhanced Amplitude, for The second-order difference of S5022, calculate the peak and trough significance, and remove the peaks whose peak and trough significance is lower than the significance threshold. The peak and trough significance calculation formula is: ， in, is the peak and trough significance, is the peak amplitude, is the trough amplitude.

9. The method for identifying railway perimeter intrusions by frequency domain enhanced vibration spectrum analysis according to claim 1 is characterized in that: The classifier adopts random forest algorithm.

Images