A city rail track damage detection method based on vibration characteristic monitoring

Abstract

The present application relates to the technical field of rail damage detection, in particular to a rail damage detection method based on vibration feature monitoring, comprising: arranging sampling points according to the topology structure of the rail, obtaining the vibration signals at the transverse, longitudinal and vertical positions of the monitored rail to form a vibration feature set; taking each sampling point as a node, extracting the vibration features of each node, and forming a probability distribution interval according to the characteristic offset of each vibration feature related to the rail damage position; recording the rail damage times corresponding to the probability distribution interval under multiple time windows, identifying the damage events corresponding to the rail damage times, and positioning the signal mutation period and the transient change amplitude; performing modal separation according to the signal mutation period and the transient change amplitude of each node, describing the damage probability of each node, and outputting the damage state classification according to the update order of the damage probability of each node; the accuracy of rail damage position positioning is realized, and the efficiency of rail damage propagation processing is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rail damage detection, in particular to a rail damage detection method based on vibration feature monitoring. BACKGROUND

[0002] In the urban rail transit system, the rail is one of the key infrastructures, and its structural integrity directly affects the safety and stability of train operation. Due to long-term bearing of high-frequency wheel-rail impact load, environmental corrosion and fatigue damage, etc., the rail is prone to structural defects such as cracks, peeling and crushing. The traditional manual inspection or defect detection vehicle detection method has the following problems: low detection frequency, difficult to real-time master the rail state; limited recognition ability for small damage; unable to predict the damage trend; lack of dynamic modeling of damage propagation path.

[0003] For example, Chinese patent publication No. CN111562307A discloses a rail damage number real-time statistical method based on direct current magnetic flux leakage principle, which belongs to the technical field of railway rail magnetic flux leakage inspection. The damage number real-time statistical method includes adaptive threshold solving, peak-to-peak value damage judgment, peak point window positioning, window characteristic value extraction and zeroing, and window cycle damage judgment counting.

[0004] For example, Chinese patent publication No. CN111896625A discloses a rail damage real-time monitoring method, which includes arranging a monitoring system on both sides of the rail; obtaining real-time monitoring data obtained by the monitoring system; performing data processing on the real-time monitoring data to obtain a processed damage picture; identifying the processed damage picture using a neural network architecture to obtain an identification result; and performing real-time warning of rail damage based on the identification result and realizing real-time monitoring of rail damage. The present application also discloses a monitoring system for realizing the above-mentioned rail damage real-time monitoring method.

[0005] The prior art respectively explains the use of peak value and peak-to-peak value to judge the cycle of rail damage, and uses time-frequency analysis and image form to obtain the realization of rail damage detection. However, in the prior art, only the peak-to-peak value can identify the signal received by the rail at the moment, and the characteristics of the peak value will be different when the distance between the sampling point and the damage position is different, which will cause the omission of some gradually changing damage positions, lack of description of the damage propagation path in the spatial position, and difficult to realize the regional classification and output of rail damage at each node. SUMMARY

[0006] In order to solve the above technical problems, the technical scheme adopted by the present application is: a rail damage detection method based on vibration feature monitoring, comprising: S1, arranging sampling points according to the topological structure of the rail, obtaining vibration signals at the transverse, longitudinal and vertical positions of the monitored rail, and forming a vibration feature set.

[0007] S2, taking each sampling point as a node, extracting the vibration features of each node, and forming a probability distribution interval according to the feature offset of each vibration feature related to the rail damage position.

[0008] S3, recording the rail damage times corresponding to the probability distribution intervals in multiple time windows, identifying the damage events corresponding to the rail damage times, and locating the signal mutation period and the transient change amplitude.

[0009] S4, modal separation according to the signal mutation period and the transient change amplitude of each node, describing the damage probability of each node, and outputting the damage state classification according to the update order of the damage probability of each node.

[0010] The beneficial effects of the present application are as follows: first, the present application arranges multiple sampling points, collects vibration signals according to the distribution position and working condition of the sampling points, and then sets up an identification based on the reference features, identifies the feature offset, generates a probability distribution interval, and preliminarily identifies the damage, so that the comprehensive expression of the vibration features contained in the vibration signal is enhanced, and the sensitivity to different types of damage is enhanced.

[0011] Second, the present application uses multiple time windows to count the changes of each data in the probability distribution interval, and judges whether the rail exists state conversion under the current measurement based on the basic event, and combines the period and instantaneous value of the damage event interval to describe the evolution of the vibration signal at each position over time, which provides a calculation basis for subsequent description of whether the damage event exists evolution and its path.

[0012] Third, the present application identifies the vibration signal according to the frequency characteristics, combines the damage probability of each node, identifies the time evolution order of the vibration signal at each position in the scene of rail sampling position and frequency level, and then generates a classification related to the damage state by identifying the propagation path with the shortest connection path, which makes up for the defects of classification based only on peak value, peak-to-peak value, effective value and other time domain statistics, improves the consistency of classification and the explainability of spatial position, and finally completes the identification and positioning of rail damage according to the output damage state classification. BRIEF DESCRIPTION OF DRAWINGS

[0013] The present application will be further described below in conjunction with the drawings and examples.

[0014] Figure 1 It is a flowchart of a rail damage detection method based on vibration feature monitoring.

[0015] Figure 2 It is a flowchart of step S2 of a rail damage detection method based on vibration feature monitoring.

[0016] Figure 3 is a flowchart of step S3 of the rail damage detection method based on vibration feature monitoring.

[0017] Figure 4 is a flowchart of step S4 of the rail damage detection method based on vibration feature monitoring. DETAILED DESCRIPTION

[0018] Embodiments of the present application are described in detail below. The embodiments described below are exemplary only, and are not to be taken in a limiting sense. Unless otherwise indicated, technical or scientific terms used in the embodiments have the same meaning as commonly understood by one of ordinary skill in the art to which the present application pertains.

[0019] Reference Figure 1 A rail damage detection method based on vibration feature monitoring includes: S1, arranging sampling points according to the topology structure of the rail, obtaining vibration signals at the transverse, longitudinal and vertical positions of the monitored rail, and forming a vibration feature set.

[0020] S2, taking each sampling point as a node, extracting the vibration feature of each node, and forming a probability distribution interval according to the characteristic offset of each vibration feature related to the rail damage position.

[0021] S3, recording the rail damage times corresponding to the probability distribution interval under multiple time windows, identifying the damage event corresponding to the rail damage times, and locating the signal mutation period and the transient change amplitude.

[0022] S4, performing modal separation according to the signal mutation period and the transient change amplitude of each node, describing the damage probability of each node, and outputting the damage state classification according to the update order of the damage probability of each node.

[0023] Preferably, the topology structure of the rail represents the structural network of the rail head, the rail alignment and the rail position on the rail, which is used to divide the rail into multiple component blocks according to the route form, so as to identify the features of the damaged part in the rail.

[0024] Preferably, the arrangement of the sampling points represents the arrangement along the longitudinal direction of the rail, focusing on covering the stress concentration areas of the rail waist and rail bottom; for identifying whether there is a change in the vibration component on the stress concentration area, the sampling points are arranged in the longitudinal direction of the rail at intervals; then a 45° inclined sensor is installed below the rail head tread to capture the tangential vibration component of the rail surface peeling; this part is used to obtain the tangential vibration caused by the rail surface peeling phenomenon, and the collected vibration signal is divided into multiple components to identify the high-frequency component and the offset part. At the same time, when arranging the sampling points, the surrounding of the damaged rail is set, and multiple groups of sampling points are arranged to identify the change form of the vibration characteristics at different distances near the damage.

[0025] The longitudinal direction refers to monitoring the tensile or compressive vibration of the rail bottom, generally representing a low-frequency vibration signal of 0-500Hz; the transverse direction refers to capturing the bending vibration of the rail waist, generally representing a medium-frequency vibration signal of 500-2000Hz; the vertical direction refers to detecting the impact vibration of the rail head, generally representing a high-frequency vibration signal of 2-10kHz; the three kinds of vibration signals are converted into vibration components in the corresponding direction according to the arrangement of the sampling points.

[0026] That is, the implementation manner of step S1 further includes: obtaining the actual working condition of the vibration signal, the actual working condition including no load, full load, and speed gradient of train running, to obtain the vibration signal under the corresponding working condition; the vibration signal at each sampling point position is counted according to the actual working condition to form a vibration feature set; and each element in the vibration feature set is sorted according to the distance of the sampling point from the damage position of the rail.

[0027] In an embodiment of the present application, as shown in Figure 2 The implementation manner of step S2 includes: S21, multi-axis signal decomposition is performed on the vibration signals at the transverse, longitudinal and vertical directions, vibration features of the vibration feature set at each direction are obtained in sequence, time sequence statistics are performed on the vibration features, and the vibration features are set with spatial identifiers according to the numerical interval of the peak value, peak-to-peak value and effective value of the vibration features after statistics. The peak value described herein represents the single-peak maximum value when the vibration signal fluctuates, the peak-to-peak value represents the difference between one peak top and the adjacent trough of the vibration waveform, to reflect the maximum change range of the signal in the time period, thereby describing the fluctuation of the current vibration signal, and the effective value represents the root mean square value of the vibration signal, used to describe the energy intensity and stability of the vibration signal in a period of time.

[0028] The spatial identifier is to associate the vibration features identified at this time with the spatial position of the collected vibration signal, to identify the value of the vibration features relative to the damage position of the rail, to identify when the vibration features are prominent at a distance from the damage position, so as to facilitate subsequent inspection of whether there is a potential damaged part in the rail.

[0029] S22, according to the spatial identification of the vibration feature, describing the spatial weight coefficient of each node to the rail damage position, and sequentially calculating the feature offset corresponding to the vibration feature and the reference feature.

[0030] S23, taking the position of each node as a prerequisite, identifying the difference amplitude of each node and the rail damage position at the corresponding distance, and combining the probability distribution interval according to the conditional probability value corresponding to the feature offset and the difference amplitude.

[0031] It should be noted that when identifying the rail damage, the vibration signals collected by each node at different distances will show different value conditions in terms of effective value, peak value, peak-to-peak value, etc. In this case, multiple sampling points are laid out to collect multiple data indicating the existence of rail damage, which may point to a position where damage exists. At this time, taking the damage at a certain position on the rail as the starting point of identification, the vibration signals near this point are identified, and the rail damage near this point is identified in combination with the distance position. According to the signal characteristics of this part of the identification, the accuracy of subsequent positioning of the damage position is improved, and the reduction of the positioning accuracy of the rail damage is reduced.

[0032] Preferably, the reference feature represents the vibration signal collected under the health condition of the train, and the average value of the vibration signal collected multiple times is selected to calculate the difference between the current vibration feature and the reference feature. The spatial weight coefficient represents the reciprocal of the distance between each node and the rail damage position. At this time, the content of the calculation represents the influence of the rail damage on the surrounding nodes. As for the feature offset, it is obtained by multiplying the difference between the current node value and the reference feature value by the spatial weight coefficient. This feature offset mainly describes the offset state of each node relative to the normal condition based on the distance from the rail damage position. The calculation result near the damage position is mainly emphasized, and the calculation value far from the damage position is reduced. Under this position-based feature offset, the calculation range for the damage position can be reduced, and other damages existing in the current rail can be identified according to the calculation content similar to the feature offset. It should be noted that this rail damage position is measured and identified after the damage is found, and the vibration signal characteristics at this position are used to check the entire rail one by one.

[0033] Preferably, the part with the highest similarity in the vibration feature of each node is also required to be output as the difference amplitude according to the position of the vibration feature sampled. That is, the implementation of the difference amplitude includes: the value, peak value, peak-to-peak value, and effective value of the corresponding time point in the vibration signal are sequentially taken as the similarity calculation content, and the value of the vibration signal is sequentially calculated with the reference feature. The similarity can be calculated in the form of Pearson correlation coefficient, and the part with the maximum similarity value is selected as the difference amplitude at this time. That is, when calculating the feature offset, the difference value of the corresponding time point value, peak value, peak-to-peak value, etc. is obtained, and then the part with the maximum similarity and the frequency of the corresponding feature offset value appearing in the historical data are selected to generate the probability distribution interval to explain that after the current node collects the vibration signal with damage problems, the data with problems is converted into a probability distribution. After mapping the offset to the probability distribution, intelligent diagnosis can be further performed using Bayesian classification, hypothesis testing, and anomaly detection algorithm. Different forms of rail damage such as fatigue cracks and peeling are distinguished.

[0034] The probability distribution interval is output as the corresponding conditional probability of each node, the distance from the rail damage position, and the value of the vibration signal mapped by the conditional probability.

[0035] When generating the probability distribution interval, the distance of each node from the rail damage position can be used as the horizontal axis, and the corresponding conditional probability of each node can be used as the vertical axis to show the distribution of related values in the probability distribution interval.

[0036] Preferably, after the vibration feature data is counted, it is also necessary to determine whether there is a feature offset in each node. If the data of these nodes shows a stable phenomenon and does not exceed the confidence interval set based on the average value of the historical data, the related nodes are removed, and the probability distribution interval is set for each value interval in the data in the value interval of the remaining nodes.

[0037] That is, the implementation of the time series statistics of the vibration feature in step S21 also includes: selecting the vibration signals at the lateral, longitudinal, and vertical directions, detecting the nodes at each direction at the same time interval, identifying the nodes with feature offset in the value interval of the vibration feature in each time interval, and performing spatial identification on the corresponding nodes.

[0038] Preferably, when the multi-axis signal is decomposed, the vibration features in three directions are separately obtained, and each vibration feature is regarded as a processing node. After comparing each node with the reference feature, the processing is completed.

[0039] In one embodiment of the present application, in step S3, according to the occurrence of the feature offset, the time period of the number of rail damage is identified, and according to the mutation part in the corresponding time period, the peak variation gradient is recorded, so as to know the change range of the damage condition represented by the rail damage at multiple positions when the rail damage exists.

[0040] When recording the number of rail damage, the value in the probability distribution interval and the duration of the value in the corresponding time period are mainly searched in combination with the characteristic value of the vibration signal, such as the peak value of the vibration characteristic, the interval of the peak value, etc., and the probability distribution interval of the vibration signal is converted into a basic event for each number of times. If there are adjacent basic events with the same time period and the search type is close, the basic events are combined to count the number of rail damage in multiple time windows, and then the events corresponding to the basic events are located to identify the period and mutation value of the vibration signal when the obvious rail damage occurs, which are output as the detection content.

[0041] Preferably, the recorded number of rail damage represents the cumulative number of damage events in which the vibration signal characteristic value deviates significantly from the healthy state reference value due to rail damage in a specific time window.

[0042] The time window described above can be set to 10 seconds, which can cover most of the time of the passing train, and a 1-second sliding step is set in the time window to make the vibration signal collected in the time window as complete as possible. Then, a plurality of short windows such as 0.5 seconds can be divided for this window, and the data in the short window can be used for transient capture of the vibration signal to identify the correlation between the collected data at each node and the rail damage.

[0043] As shown in Figure 3 The implementation of step S3 includes: S31, time searching is performed with the mapping relationship between the probability distribution interval and each node to determine the duration of the corresponding node, and the number of basic events corresponding to each node is queried by combining the values of each node in the probability distribution interval in multiple time windows. Here, the number of basic events queried is taken as the number of rail damage according to the damage events existing in the basic events.

[0044] S32, it is judged whether each basic event has state conversion, and the damage events in each basic event are extracted according to the state conversion identification result of each basic event, and the number of rail damage is set according to the number of damage events in the corresponding time window.

[0045] S33, traversing each damage event with a time window, locating the period length of the vibration signal under the corresponding rail damage frequency according to the mapping relationship between the damage event and the probability distribution interval, outputting the signal mutation period as the output, and outputting the instantaneous value corresponding to the signal mutation period as the transient change amplitude.

[0046] The state transition indicates that the node in a certain state enters another state, such as the vibration signal in the abnormal state entering the abnormal state, or in the identified time window, one time window approaches mild damage, and one approaches moderate damage, indicating that the current identified position may have a more serious damage condition, resulting in obvious fluctuations in the data.

[0047] As shown in Table 1, taking one of the cases existing in the queried damage event as an example, and the duration of the vibration signal corresponding to the conditional probability of each node, the damage level about the abnormal condition existing in the basic event is queried.

[0048] Table 1. Damage diagram

[0049]

[0050] In Table 1, the results of searching based on the abnormal conditions existing in part of the basic events are illustrated, which represent the range of characteristic values of the vibration characteristics and the corresponding probability threshold when the rail is damaged; At the same time, the above-mentioned probability threshold indicates that the value of the node in the corresponding distance interval approaches the corresponding multiple of the reference value, and the duration and the probability value of the node that meets the approach are added to meet the probability threshold range, indicating that there is a corresponding damage level at the position represented by the corresponding distance interval. As illustrated in the first row, the frequency of damage existing in the historical data under a certain distance from the rail damage position and its characteristic offset is only when the identified data meet the duration and the ratio respectively, it will be considered that there is a corresponding damage level; The peak ratio, the peak-to-peak ratio, and the effective value ratio are part of the proportion value used by the current vibration characteristics and the reference characteristics; It should be noted that the damage content appearing here not only includes the content shown in the table, but also other forms of rail damage, which is only to illustrate that when the damage event is searched, the corresponding data may exist.

[0051] As shown in Table 2, the relative conditions of the peak value, the peak-to-peak value, and the effective value that may be collected under some basic events are illustrated.

[0052] Table 2. Event diagram

[0053]

[0054] In Table 2, in addition to event 1, other events represent the existence of injury content. Here, the peak value, peak-to-peak value and effective value under the existence of injury are displayed to illustrate the form of vibration signal recognition when injury occurs. In Table 2, event 2 illustrates the existence of injury. As for events 3 to 5, the distance from the rail injury position is sorted from near to far to illustrate the existence of injury. Then, the values of this part are compared with the values of the historical data benchmark characteristics to determine how close the values are to the benchmark values. After that, the existence condition probability is compared with the condition probability of the historical data to obtain the probability interval range, and the time window of the divided basic events is identified to set the duration. Finally, the distance from the rail injury position of the obtained data is described to describe the content of the current basic event.

[0055] In step S31, the time retrieval mapping is used to identify and calculate the time interval corresponding to each probability value in the probability distribution interval, and the time period in which the injury or no injury is continuously represented in these time intervals is taken as the duration to query the possible basic events of each node. The basic events include normal vibration signals in the normal confidence interval and damage events with injury. The damage events are further divided into multiple types to describe the possible problems of rail injury.

[0056] When identifying whether each basic event exists in the state transition, the continuous basic events are used for judgment, the types of each basic event are identified, and the rail injury of each node in the current probability distribution interval is viewed in a dual-channel manner. After quickly detecting obvious injury, the gradual damage characteristics are captured through time-frequency analysis or state transition comparison, and then the existing damage events are counted by using the sliding window. At the same time, whether the state transition part is normal is judged by combining the characteristic values of the vibration energy and frequency domain characteristics of the vibration signal to obtain the corresponding number of rail injuries.

[0057] The implementation of step S32 also includes double-channel identification of each basic event to obtain its preliminary state identification and progressive state identification in a continuous time window, respectively. The state identification is the content of identifying the basic event when searching for the event in the historical data in terms of the probability value, characteristic value, duration, etc. of the vibration signal. The basic event at this time represents the vibration signal collected by each node under a single time window. If the vibration signal is within the normal value range, it is a non-damaged basic event. If it is not within the normal range, the relevant value is searched and identified to determine what kind of damage event the basic event represented by the current vibration signal belongs to, such as a slight, moderate, or severe damage event, and the damage event is counted and its specific data situation is explained according to its time period and instantaneous value. In addition to identifying the probability value, the harmonic proportion and high-frequency energy ratio of the vibration signal can also be used to identify the state identification of the current basic event.

[0058] The preliminary state identification and the progressive state identification described herein serve as the state identification result to complete the identification of each basic event.

[0059] The preliminary state identification is used to identify the basic event after identification, and based on the vibration signal characteristics within a single time window, it is directly classified as a non-damaged or damage event. For example, a sudden increase in the harmonic proportion leads to the identification of surface cracks under high-frequency vibration; a deviation of the probability value from the baseline leads to the identification of loose bolts under low-frequency vibration amplitude anomalies; and a significant increase in the high-frequency energy ratio leads to the identification of external impact on the rail.

[0060] The progressive state identification then tracks the long-term trend of the vibration signal through cumulative analysis of consecutive time windows, such as the rate of change of the characteristic value over time and the count and distribution of damage events within the time window, and then classifies the damage events into multiple severity levels.

[0061] The preliminary state identification and the progressive state identification are combined according to a preset rule, and each damage event is marked with a damage level according to the preset rule combination.

[0062] The preset rule indicates the damage event that can be satisfied after combining the preliminary state identifier and the progressive state identifier identified by the single node, and the preset rule is used to describe the meaning of the combination of the plurality of state identifiers, such as that after acquiring a moderate damage event in three continuous time windows, the combination can be identified as severe, that is, the content identified in the data set after combining the preliminary state identifier and the progressive state identifier, describing the severity, and the severity of the preliminary state identifier and the progressive state identifier can also be set with a score, and the weighted sum thereof is used as the content contained in the preset rule, such as that the weights of the preliminary state identifier and the progressive state identifier are set as 0.6 and 0.4, respectively, to weight them and identify the severity of the corresponding damage event.

[0063] The preliminary state identifier is described based on the conditional probability of the feature offset amount in the probability distribution interval, and when the feature value corresponding to the conditional probability exceeds the confidence interval, the time window in which the node is located is directly marked as a damage event, and then the difference amplitude is used to further describe the severity of the node corresponding to the preliminary state identifier, and the preliminary state identifier has a score value in the range of 0 to 1 after being acquired, which is obtained by querying the database after determining the preliminary state identifier.

[0064] The progressive state identifier uses the change rate of the conditional probability, the special offset amount and the difference amplitude in the probability distribution interval in the continuous time window to identify the trend, and uses the change rate of the three values to search in the database to find the progressive state identifier of the current node; and the score value of the queried progressive state identifier can be used to further describe the severity of the current damage event.

[0065] Preferably, the signal mutation period described in step S33 is used to describe the interval period of the corresponding data identified under a certain number of rail damage events, and the transient change amplitude is used to describe the instantaneous value in the interval period. The implementation of the signal mutation period also includes: for the same type of damage event, calculating the time interval of adjacent events after time sorting; and then selecting the time interval of the current damage event as the output signal mutation period according to the mode of the time interval of the adjacent events.

[0066] The transient change amplitude is the maximum instantaneous amplitude in the signal mutation period minus the baseline amplitude, and the amplitude is the amplitude value of the fluctuation of the vibration signal; the baseline amplitude is generally the average value of 1 second before the period.

[0067] In one embodiment of the present application, when modal separation is performed, the data in the low, high and medium frequency states is segmented, and damage mode matching is performed based on the identified damage events. Then, the position of each node and the order of updating are introduced, the damage propagation path in the rail is considered, and each node is classified and output to represent the classification of the damage state.

[0068] As shown in Figure 4 The implementation mode of step S4 includes: S41, segmenting the data of each node in the signal mutation period according to the frequency characteristics to obtain the modal feature set after modal separation; at this time, the frequency characteristics represent that the vibration signals in the transverse, longitudinal and vertical directions are mapped with the signal mutation period, the vibration signals in various different collection modes are found out, the position of the identified damage event is further shortened, and the more accurate rail damage position is positioned.

[0069] S42, damage mode matching is performed on the damage events included in the modal feature set to obtain the damage probability corresponding to each node; at this time, the damage mode matching is performed by calculating the similarity between the damage events and the existing damage types in the damage mode library, and the similarity is calculated by using the peak value, peak-to-peak value, effective value and other data extracted from the vibration signals. When calculating the similarity, the calculation method of Pearson correlation coefficient is adopted, the average value of the Pearson correlation coefficients calculated from these values is taken as the calculated similarity, and the part with the maximum similarity value is taken as the corresponding damage mode. Then, the Gaussian distribution form in the Gaussian mixture model is used according to the values corresponding to the damage events, the values of the vibration signals bound to the damage events are taken as inputs, the values, average values and standard deviations of each time point are used to complete the calculation of Gaussian distribution, and the probability value calculated by Gaussian distribution is taken as the damage probability of each node.

[0070] S43, sort each damage event according to the order of occurrence time to form the time evolution order of each damage event, and obtain the shortest connection path of the damage event as the propagation path according to the time evolution order.

[0071] At this time, the shortest connection path is the path that connects to each damage event on the time evolution order as much as possible. This shortest connection path can use Dijkstra algorithm, take the order of occurrence time as the connection direction, take the damage probability as the weight, take the value of the damage event in the probability distribution interval as the value when calculating the shortest connection path, and take the weight corresponding to this value as the way to identify the shortest path. The nodes represented by each damage event are connected to obtain the shortest connection path to represent the trend of damage propagation existing in the current rail.

[0072] S44, sequentially classify each node according to the propagation path, and output the classified nodes as the damage state classification.

[0073] The classification of each node in the propagation path is performed by dividing the nodes existing in the propagation path into multiple clusters according to their damage probabilities and positions in the propagation path. At this time, the K-means clustering method is used to classify the previously processed data for each node, describe whether the vibration signal at different positions exists in the propagation process, and materialize these processes to represent the multiple damage conditions identified in the current rail. This damage state classification can directly support the integration of maintenance decision and fault warning systems, facilitate the understanding of the rail damage development mechanism, and facilitate the adjustment of the priority of subsequent rail maintenance to improve the efficiency of rail maintenance decision-making.

[0074] Meanwhile, there are multiple data types in the damage state classification, and the update order needs to be verified. The damage state classification is adjusted according to the classification of each node in the update order. For the update order, the updated time point is used as the sequence to make up for the multiple unstable damage events ignored according to the time sequence of the occurrence of each damage event.

[0075] Preferably, when classifying the damage probability and the position in the propagation path, the damage probability is used as the value for classification, the ratio of the distance of each node in the propagation path to the total distance of the propagation path is used as the weight, and the K-means clustering method is used to generate multiple clusters to complete the classification of each node. After adding the update order, the normalized value is multiplied by the current weight to adjust the weight and the cluster, thereby completing the classification of the damage state of each node.

[0076] In step S44, when the classified nodes are output as the damage state classification, the implementation method further includes adjusting the classified nodes according to the update order of each node, aggregating the classified nodes using the time interval of the update order, and outputting the aggregated nodes as the damage state classification.

[0077] At this time, after normalizing the time interval according to the update order, the normalized value is used as the weight, and the damage probability is used as the calculation value. The weighted average of each node in the cluster is calculated, and each cluster is adjusted until the average value of the data in each cluster is not changed. The data of the corresponding cluster is output as the damage state classification.

[0078] Although the embodiments of the present application have been shown and described above, it is to be understood that the above-described embodiments are merely exemplary, and are not to be construed as limiting the present application, and that changes, modifications, substitutions and variations can be made thereto without departing from the scope of the present application.

Claims

1. A method for detecting rail damage of urban rail based on vibration characteristic monitoring, characterized in that, The method comprises the following steps: S1, arranging sampling points according to the topology of the steel rail, obtaining vibration signals at the transverse, longitudinal and vertical positions of the monitored steel rail, and forming a vibration feature set; S2, taking each sampling point as a node, extracting the vibration features of each node, and forming a probability distribution interval according to the characteristic offset of each vibration feature relative to the position of the rail damage; S3, recording the number of rail damages corresponding to the probability distribution interval in multiple time windows, identifying the damage events corresponding to the number of rail damages, and locating the signal mutation period and the transient change amplitude; S4, modal separation according to the signal mutation period and the transient change amplitude of each node, describing the damage probability of each node, and outputting the damage state classification according to the update order of the damage probability of each node; The implementation of step S2 comprises: S21, multi-axis signal decomposition of the vibration signals at the transverse, longitudinal and vertical positions, sequentially obtaining the vibration features of the vibration feature set at each direction, time series statistics of the vibration features, and setting the spatial identifier of the vibration features according to the numerical interval of the peak value, peak-to-peak value and effective value of the statistical vibration features; S22, according to the spatial identifier of the vibration features, describing the spatial weight coefficient of each node to the position of the rail damage, and sequentially calculating the characteristic offset corresponding to the reference feature; S23, taking the position of each node as a prerequisite, identifying the difference amplitude of each node and the position of the rail damage at the corresponding distance, and combining the conditional probability values corresponding to the characteristic offset and the difference amplitude to form the probability distribution interval; The implementation of step S3 comprises: S31, time retrieval according to the mapping relationship between the probability distribution interval and each node, determining the duration of the corresponding node, and querying the basic event number corresponding to each node according to the values of each node in multiple time windows; S32, judging whether each basic event has state conversion, and extracting the damage event in each basic event according to the state conversion identification result of each basic event, and setting the number of rail damages according to the number of damage events in the corresponding time window; S33, traversing each damage event according to the time window, locating the period length of the vibration signal of the damage event at the corresponding number of rail damages according to the mapping relationship between the damage event and the probability distribution interval, and outputting the period length as the signal mutation period, and outputting the instantaneous value corresponding to the signal mutation period as the transient change amplitude.

2. The rail damage detection method based on vibration feature monitoring according to claim 1, characterized in that, The implementation of step S1 further comprises: Obtaining the actual working condition of the vibration signal, the actual working condition including the speed gradient of the empty load, the full load and the train running, so as to obtain the vibration signal under the corresponding working condition; statistically obtaining the vibration signal at each sampling point according to the actual working condition to form the vibration feature set; and sorting each element in the vibration feature set according to the distance between the sampling point and the position of the rail damage.

3. The rail damage detection method based on vibration feature monitoring according to claim 1, characterized in that, The implementation of time series statistics of the vibration features in step S21 further comprises: Selecting the vibration signals at the transverse, longitudinal and vertical positions, detecting the nodes at each direction at the same time interval, identifying the nodes with characteristic offset according to the numerical interval of the vibration features in each time interval, and spatially identifying the corresponding nodes.

4. The rail damage detection method based on vibration feature monitoring according to claim 1, characterized in that, The implementation of the difference amplitude comprises: The values, peak values, peak-to-peak values and effective value parts of the corresponding time points in the vibration signal are sequentially taken as the similarity calculation contents, the values of the vibration signal and the reference features are sequentially calculated for similarity, and the part with the maximum similarity value is selected as the difference amplitude at this time.

5. The rail damage detection method based on vibration feature monitoring according to claim 1, characterized in that, The implementation manner of step S32 further includes: The basic events are identified in two channels to obtain their preliminary state identifiers and progressive state identifiers in continuous time windows, respectively; The preliminary state identifiers and the progressive state identifiers are combined according to a preset rule, and the damage level under the preset rule combination is marked for each damage event.

6. The rail damage detection method based on vibration signature monitoring according to claim 4, characterized in that, The implementation manner of the signal mutation period further includes: For damage events of the same type, the time interval of adjacent events after time sorting is calculated; and according to the mode of the time interval of adjacent events, the time interval of the current damage event is selected as the output signal mutation period.

7. The rail damage detection method based on vibration feature monitoring according to claim 1, characterized in that, The implementation manner of step S4 includes: S41, the data of each node under the signal mutation period is segmented according to the frequency characteristics to obtain the modal feature set after modal separation; S42, the damage events contained in the modal feature set are matched for damage mode to obtain the damage probability corresponding to each node; S43, each damage event is sorted according to the appearance time to form the time evolution order of each damage event, and the shortest connection path of the damage event is obtained as the propagation path according to the time evolution order; S44, each node is classified according to the propagation path, and the classified nodes are taken as the damage state classification output.

8. The rail damage detection method based on vibration feature monitoring according to claim 7, characterized in that, When the classified nodes are taken as the damage state classification output in step S44, the implementation manner further includes: The classified nodes are adjusted according to the update order of each node, the classified nodes are aggregated by using the time interval of the update order, and the aggregated nodes are taken as the damage state classification output.

Images