Railway axle counting equipment online fault diagnosis method and system

Abstract

The present application relates to the technical field of fault prediction, in particular to a railway axle counting equipment online fault diagnosis method and system, comprising the following steps: collecting original analog voltage signals of the axle counting sensor in real time, sampling each wheel passing signal to obtain original voltage signal samples, and performing time axis stretching and contraction transformation on each original voltage signal sample to establish a standardized waveform sequence. The present application improves the comparability of the waveform in the time dimension by sequentially sampling and standardizing the original analog voltage signals of the axle counting sensor, and constructs a microscopic morphological feature vector combining multiple waveform characteristics such as peak symmetry, rising edge slope, falling edge slope, post-zero-point oscillation decay rate and integral area, so that the fine-grained information reflecting the health status of the sensor can be modeled.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault prediction, in particular to a railway axle counting equipment online fault diagnosis method and system. BACKGROUND

[0002] The railway axle counting equipment online fault diagnosis method is used for monitoring and abnormity judgment on the health state of key equipment, axle counting sensors in a railway signal system. Although the prior art has the online monitoring capability on the health state of the axle counting sensor, in actual application, it depends on the single-cycle signal feature for fault judgment, and it is difficult to identify the slow-changing performance degradation process, for example, in the case that the internal coil of the sensor gradually loosens or the magnetic field response weakens, the single-cycle voltage signal may still be approximately normal, which makes the prior art difficult to capture potential hidden dangers in time, and only when the abnormality is serious and accumulates can it trigger a response, increasing the risk of fault outbreak. Therefore, improvement is needed. SUMMARY

[0003] The purpose of the present application is to solve the shortcomings in the prior art and provide a railway axle counting equipment online fault diagnosis method and system.

[0004] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme, a railway axle counting equipment online fault diagnosis method, comprising the following steps:

[0005] Online real-time acquisition of the original analog voltage signal of the axle counting sensor, sampling of each wheel passing signal to obtain the original voltage signal sample, time axis stretching and transformation of each original voltage signal sample to establish a standardized waveform sequence;

[0006] According to the standardized waveform sequence, extracting parameters of each standardized waveform in the sequence to generate a multi-dimensional feature parameter group, and combining each multi-dimensional feature parameter group associated with the sensor coil state into a microscopic morphological feature vector set;

[0007] According to the microscopic morphological feature vector set, sequentially taking out two continuous feature vectors to calculate the Fréchet distance between the two feature vectors to obtain an instantaneous morphological similarity, and operating the instantaneous morphological similarity with the running total score of the last cycle to obtain a reflection consistency score;

[0008] According to the reflection consistency score, tracking the numerical change in continuous multiple cycles to determine whether the numerical value presents a monotone decreasing trend to generate a score degradation trend judgment result, comparing the latest reflection consistency score with a preset warning threshold, combining the score degradation trend judgment result, and outputting a slow-changing fault warning state of the axle counting equipment.

[0009] Preferably, the step of obtaining the standardized waveform sequence is:

[0010] According to the original analog voltage signal of the axle counting sensor, the start and end time indexes of each wheel passing signal are located, the continuous voltage value segment is intercepted according to the start and end time indexes, the voltage value is recorded according to the fixed sampling interval, and the original voltage signal sample is obtained;

[0011] According to the original voltage signal sample, the proportion of each time span to the target time length is calculated, the time axis is stretched or compressed according to the proportion, the peak position sequence and zero position are kept consistent according to the uniform time step, and the voltage value sequence after time axis stretching is generated;

[0012] According to the voltage value sequence after time axis stretching, the equal length sample from the uniform start point to the uniform end point is intercepted, the sequence order is arranged according to the uniform sample number and the start point zero value is aligned, and the standardized waveform sequence is formed.

[0013] Preferably, the acquisition step of the multi-dimensional feature parameter group is:

[0014] According to the standardized waveform sequence, the amplitude of each standardized waveform is read, the peak sample position is located, the zero-crossing point position is located, the starting point of the rising edge and the termination point of the falling edge are labeled, and the waveform structure labeling result is obtained;

[0015] According to the waveform structure labeling result, the amplitude difference is counted in the left and right intervals of the peak sample position, the waveform peak symmetry is formed, the rising edge slope is formed according to the amplitude increment and time span from the starting point of the rising edge to the peak sample position, the falling edge slope is formed according to the amplitude decrement and time span from the peak sample position to the termination point of the falling edge, the interval is divided by a fixed step after the zero-crossing point, the ratio of adjacent amplitudes is compared segment by segment, the oscillation decay rate is formed, and the waveform integral area is formed by point-by-point accumulation of all amplitudes under the uniform time step, and the multi-dimensional feature parameter group is generated.

[0016] Preferably, the acquisition step of the micro-morphology feature vector set is:

[0017] According to the multi-dimensional feature parameter group, the sensor coil state identifier corresponding to each standardized waveform is read, the sensor coil state identifier is appended to the end of the multi-dimensional feature parameter group, and the micro-morphology feature vector set is formed.

[0018] Preferably, the acquisition step of the instantaneous morphology similarity is:

[0019] According to the micro-morphology feature vector set, the adjacent two feature vectors are extracted by sliding one bit in time sequence, aligned one by one according to the component index, the missing components are filled by interpolation with adjacent effective components, and the pair order and time identifier are recorded consistent, and the adjacent feature vector pair sequence is obtained;

[0020] According to the adjacent feature vector pair sequence, a component-to-component monotone matching path is established pair by pair, a Frechet distance numerical value of each pair is calculated, the Frechet distance numerical value is normalized into a similarity numerical value by a linear mapping of a zero to a one interval, and an instantaneous shape similarity is obtained.

[0021] Preferably, the acquisition step of the reflection consistency score is:

[0022] According to the instantaneous shape similarity, a reflection consistency score is calculated in combination with a running total score of a previous period.

[0023] Preferably, the acquisition step of the score deterioration trend determination result is:

[0024] According to the reflection consistency score, reflection consistency scores of consecutive multiple periods are sequentially aggregated, missing values are removed and boundary vacancies are linearly interpolated, and a continuous period reflection consistency score sequence is formed;

[0025] According to the continuous period reflection consistency score sequence, a weighted trend coefficient is calculated.

[0026] According to the weighted trend coefficient, a deterioration trend threshold value is set and compared with the weighted trend coefficient, if the weighted trend coefficient is less than the deterioration trend threshold value, it is marked as deterioration, otherwise it is marked as non-deterioration, and a score deterioration trend determination result is generated.

[0027] Preferably, the acquisition step of the slow-changing fault early warning state is:

[0028] According to the reflection consistency score, the latest reflection consistency score is read, a warning threshold value is called for single comparison, the relationship is marked as less than, equal to or greater than, the difference between the latest reflection consistency score and the warning threshold value is calculated, the comparison time stamp and the relationship label are recorded, and a reflection consistency score threshold value comparison result is obtained.

[0029] According to the reflection consistency score threshold value comparison result, the score deterioration trend determination result is read, if the relationship label is less than and the score deterioration trend determination result is yes, it is set as triggered, if the relationship label is less than and the score deterioration trend determination result is no, it is set as observed, and if the relationship label is equal to or greater than, it is set as not triggered, and a warning trigger decision result is generated.

[0030] According to the warning trigger decision result, the trigger is mapped to the slow-changing fault early warning state of the axle counting equipment as early warning, the observation is mapped to the slow-changing fault early warning state of the axle counting equipment as observation, and the non-trigger is mapped to the slow-changing fault early warning state of the axle counting equipment as normal, and the slow-changing fault early warning state of the axle counting equipment is output.

[0031] The application also provides an online fault diagnosis system, comprising:

[0032] The acquisition module is used for acquiring original analog voltage signals of the axle counting sensor in real time, sampling each wheel passing signal, obtaining original voltage signal samples, and performing time axis stretching and transformation on each original voltage signal sample to establish a standardized waveform sequence.

[0033] The feature extraction module is used for extracting parameters of each standardized waveform in the sequence according to the standardized waveform sequence, generating a multi-dimensional feature parameter group, and combining each multi-dimensional feature parameter group associated with a sensor coil state into a micro-morphology feature vector set.

[0034] The similarity calculation module is used for sequentially taking out two continuous feature vectors according to the micro-morphology feature vector set, calculating the Frechet distance between the two feature vectors to obtain an instantaneous morphology similarity, and performing operation on the instantaneous morphology similarity and an operation total score of a previous period to obtain a reflection consistency score.

[0035] The early warning judgment module is used for tracking numerical changes in continuous multiple periods according to the reflection consistency score, judging whether the numerical value presents a monotone decreasing trend, generating a score degradation trend judgment result, comparing the latest reflection consistency score with a preset early warning threshold, combining the score degradation trend judgment result, and outputting a slow change fault early warning state of the axle counting device.

[0036] Compared with the prior art, the advantages and positive effects of the present application are that:

[0037] The present application improves the comparability of the waveform in the time dimension by sequentially sampling and standardizing the original analog voltage signals of the axle counting sensor, constructs a micro-morphology feature vector by combining multiple waveform features such as peak symmetry, rising edge slope, falling edge slope, oscillation decay rate after zero crossing point and integral area, models fine-grained information reflecting the health state of the sensor, measures the difference between two vectors in each period by the Frechet distance, constructs a reflection consistency score by combining the historical operation score, gradually suppresses the fluctuations within the period in the process of continuous updating, improves the trend recognition ability, dynamically fuses the single-period fluctuations and cross-period trends, and alleviates the hypersensitive response problem of traditional single-point abnormality discrimination to sudden noise, can smoothly capture the sensor degradation process, and improves the recognition ability of latent faults. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The present application is a step schematic diagram. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0040] Referring to Figure 1 The present application provides a technical scheme, an online fault diagnosis method of a railway axle counting device, comprising the following steps:

[0041] The original analog voltage signal of the axle counting sensor is collected in real time online, each wheel passing signal is sampled, the original voltage signal sample is obtained, and each original voltage signal sample is subjected to time axis stretching and compression transformation to establish a standardized waveform sequence.

[0042] According to the standardized waveform sequence, parameters of each standardized waveform in the sequence are extracted to generate a multi-dimensional feature parameter group, and each multi-dimensional feature parameter group associated with the sensor coil state is combined into a microscopic morphological feature vector set.

[0043] According to the microscopic morphological feature vector set, two continuous feature vectors are sequentially taken out, the Fréchet distance between the two feature vectors is calculated to obtain an instantaneous morphological similarity, the instantaneous morphological similarity is operated with the running total score of the last period to obtain a reflection consistency score.

[0044] According to the reflection consistency score, the numerical value change in the continuous multiple periods is tracked, it is determined whether the numerical value presents a monotonic decreasing trend, a score degradation trend determination result is generated, the latest reflection consistency score is compared with a preset warning threshold, and the score degradation trend determination result is combined to output a slow-changing fault warning state of the axle counting device.

[0045] The acquisition step of the standardized waveform sequence is:

[0046] According to the original analog voltage signal of the axle counting sensor, the start and end time indexes of each wheel passing signal are located, the continuous voltage value segment is intercepted according to the start and end time indexes, the voltage value is recorded according to the fixed sampling interval, and the original voltage signal sample is obtained.

[0047] According to the original voltage signal sample, the proportion of each time span to the target time length is calculated, the time axis is stretched or compressed according to the proportion, the voltage value sequence after time axis stretching is resampled according to the uniform time step to keep the peak position order and zero position consistent.

[0048] According to the voltage value sequence after time axis stretching, the equal length sample from the uniform start point to the uniform end point is intercepted, the sequence order is arranged according to the uniform sample number and the start point zero value is aligned to form the standardized waveform sequence.

[0049] Specifically, according to the original analog voltage signal of the axle sensor, firstly, the continuous voltage data stream is analyzed in real time through a sliding window mechanism, a window width, for example, 50 milliseconds, is set to capture the typical duration of wheel passing, in each window, the root mean square of signal amplitude is calculated to quantify the energy of the signal, in order to accurately locate the wheel passing signal, a dynamic trigger threshold needs to be established, the setting process of the threshold is as follows, first, collect a background noise signal during a period without train passing, for example, collect a signal with a duration of 5 seconds, calculate the average value and standard deviation of the signal, for example, the measured average value is 0.03V, the standard deviation is 0.01V, then introduce a noise suppression coefficient, which is a pre-set empirical value, for example, set to 6, the selection of this coefficient is to balance between effectively filtering out background noise and avoiding missing real signals, then the dynamic trigger threshold is calculated as the average value of background noise plus the product of noise suppression coefficient and background noise standard deviation, that is, 0.03V + 6 * 0.01V = 0.09V, when the calculated signal energy in the sliding window first exceeds the dynamic trigger threshold of 0.09V, the starting time index of the signal is recorded as the previous sampling point of the time point, then the signal is continuously monitored, when the signal amplitude falls below the threshold and continuously maintains a pre-set stable duration, for example, 10 milliseconds, without exceeding the threshold again, the starting point of the stable duration is recorded as the termination time index of the signal, according to the pair of starting and ending time indexes determined in this way, a continuous voltage value segment representing a single wheel passing is accurately cut out from the original analog voltage signal stream, and according to the inherent sampling frequency of the device, for example, 20000 times per second, the cut voltage value is converted into a discrete numerical sequence, which constitutes an original voltage signal sample.

[0050] According to the original voltage signal samples, first need to set a standardization of the target length of time, the length of the determination is based on the statistical analysis of historical wheel through signal samples, for example, analysis of the past 10000 original voltage signal samples of the duration, calculate the average of all sample time length, for example, this average is 25 milliseconds, corresponding to the sampling rate of 20000 Hz for the number of points in 500 points, then the length of this 500 sampling points set as the target length of time, then for each new original voltage signal samples, calculate its own time span and the ratio of this target length of time, for example, a sample of the actual length of 600 sampling points, the scaling ratio is 500 divided by 600, equal to about 0.833, another sample length is 400 sampling points, the scaling ratio is 500 divided by 400, equal to 1.25, then, according to the calculation of the specific proportion, adjust the time axis of each sample, the adjustment process using cubic spline interpolation algorithm to execute, the algorithm according to the original sample of discrete voltage data points to generate a smooth continuous function, so as to better maintain the key features of the original waveform when resampling, such as the position of the peak value and the relative order of zero crossing, after generating a continuous function, according to the uniform time step, that is, the 500 equally spaced time points corresponding to the target length of time, the function is resampled, so, whether the original sample is longer than or shorter than the target length of time, will be unified into a new sequence containing 500 voltage value points, so as to generate the time axis after the voltage value sequence of the time length of the unified time length.

[0051] According to the time axis after the voltage value sequence is stretched, although all sequences have the same total length, for example, 500 data points, in order to eliminate the edge effect that may be introduced by the slight jitter of signal trigger or interpolation process at both ends of the sequence, further truncation is needed, by analyzing a large number of time axis after the voltage value sequence is stretched, determine the core area where the signal energy is most concentrated, for example, it is found that most of the effective waveform features appear between the 20th sampling point and the 480th sampling point, so set the unified starting point to the 20th point and the unified end point to the 480th point, truncate all sequences to get equal length samples with a length of 461 data points, then arrange these truncated equal length samples according to their original time sequence to form a data matrix, where each row represents the waveform data of a wheel passing event, finally, in order to eliminate the DC bias difference that may exist in different sensor channels or the same channel at different times, perform starting point zero value alignment correction, the specific operation is to read the voltage value of the first data point of each equal length sample, and subtract all data points in this sample from this initial voltage value, for example, the starting voltage value of a sample is [0.021, 0.035, 0.058,...], then subtract all the values in the sequence from 0.021 to become [0.000, 0.014, 0.037,...], through this operation, it is ensured that each processed waveform starts from zero voltage value, and finally a standardized waveform sequence is formed.

[0052] The acquisition step of the multi-dimensional feature parameter group is:

[0053] According to the standardized waveform sequence, read the amplitude of each standardized waveform, locate the peak sample position, locate the zero crossing point position, and mark the starting point of the rising edge and the termination point of the falling edge at the same time, to get the waveform structure annotation result;

[0054] According to the waveform structure annotation result, the amplitude difference is counted in the left and right intervals of the peak value sample position respectively to form the waveform peak value symmetry, then the amplitude increment and time span from the starting point of the rising edge to the peak value sample position are formed to form the rising edge slope, then the amplitude decrement and time span from the peak value sample position to the termination point of the falling edge are formed to form the falling edge slope, at the same time, the interval is divided by fixed step length after the zero crossing point, the ratio of adjacent amplitudes is compared segment by segment to form the oscillation decay rate, and the waveform integral area is formed by point by point accumulation of all amplitudes under the unified time step to generate the multi-dimensional feature parameter group.

[0055] Specifically, according to the standardized waveform sequence, each standardized waveform data is traversed, first, by searching the amplitude of the entire waveform sequence, the data point with the maximum absolute value is determined, and the time index of the data point is recorded as the peak sample position, then from the peak sample position, search in the starting and ending directions of the sequence respectively, locate the position where the sign changes, that is, the position where the product of the amplitude of the previous sampling point and the current sampling point is negative or zero, mark the two points closest to the peak sample position as the pre-peak zero-crossing point and the post-peak zero-crossing point respectively, then, in order to define the range of the effective signal, it is necessary to set an edge detection threshold, which is obtained by calculating a certain percentage of the peak value amplitude of the current waveform, for example, set to 8% of the peak value amplitude, this percentage is an empirical value determined based on statistical analysis of thousands of waveform samples in the historical normal running state, to filter out the small noise fluctuations at the bottom of the waveform, from the peak sample position, backtrack in the negative direction of the time axis, find the first sampling point with an amplitude lower than the dynamic edge detection threshold, mark its position as the starting point of the rising edge, similarly, search from the peak sample position in the positive direction of the time axis, find the first sampling point with an amplitude lower than the dynamic edge detection threshold, mark its position as the termination point of the falling edge, integrate the time index and amplitude information of the above located peak sample position, zero-crossing point position, starting point of the rising edge and termination point of the falling edge to obtain the waveform structure annotation result.

[0056] According to the waveform structure annotation result, first, the waveform peak symmetry is calculated. Specifically, taking the peak sample position as the center, a fixed sample number interval is taken on the left and right, for example, 40 sample points are taken, then the absolute difference of the amplitudes of the two sample points at the symmetric position is calculated point by point, for example, the difference between the amplitude at the peak sample position plus 1 and the amplitude at the peak sample position minus 1 is calculated, then the difference between plus 2 and minus 2 is calculated, and so on. Finally, the 40 difference values are added up to obtain a total sum, which is the quantization value of the waveform peak symmetry. Then, the starting point of the rising edge and the peak sample position in the waveform structure annotation result are used to calculate the amplitude increment between the two points, and then divided by the time span between the two points, which is the number of sample points multiplied by the sampling time interval, to obtain the rising edge slope. Similarly, the peak sample position and the termination point of the falling edge are used to calculate the amplitude decrement between the two points, and then divided by the corresponding time span to obtain the falling edge slope. Then, in order to calculate the oscillation decay rate, the peak value amplitude of the first and second oscillation waves following the peak value amplitude is found from the peak after zero crossing point in the waveform structure annotation result, and the ratio of the second peak value amplitude to the first peak value amplitude is obtained, which is the oscillation decay rate. Finally, the absolute values of the amplitudes of all sample points in the entire standardized waveform sequence are taken and then added up, and then multiplied by a uniform time step to calculate the waveform integral area. The five parameters of waveform peak symmetry, rising edge slope, falling edge slope, oscillation decay rate and waveform integral area are combined to form a multi-dimensional feature parameter group.

[0057] The acquisition step of the micro-morphology feature vector set is:

[0058] According to the multi-dimensional feature parameter group, the sensor coil state identifier corresponding to each standardized waveform is read, and the sensor coil state identifier is attached to the end of the multi-dimensional feature parameter group to form a micro-morphology feature vector set.

[0059] Specifically, according to the multi-dimensional feature parameter group, it is necessary to associate each parameter group with its source information, that is, which sensor coil it is generated by. In the axle counting system, at least two sensor coils are included, for example, coils A and B, which are used to detect the direction and speed of the wheel passing, so when the original analog voltage signal is collected, it is clear that each signal segment comes from a specific coil. This source information, that is, the sensor coil state identifier, will be accompanied by signal data through all preprocessing steps. Here, this identifier is digitally encoded, for example, the identifier of coil A is set to 0 and the identifier of coil B is set to 1. Then, for the multi-dimensional feature parameter group calculated from a certain standardized waveform of coil A, for example, [0.85, 15.2, -14.9, 0.3, 40.1], the code 0 representing coil A is appended to the end of the array, making it [0.85, 15.2, -14.9, 0.3, 40.1, 0]. Similarly, the same operation is performed on the multi-dimensional feature parameter group generated by coil B, and the code 1 representing coil B is appended. By performing this appending operation on all generated multi-dimensional feature parameter groups, discrete physical features are bound to their physical sources to form a micro-morphological feature vector set.

[0060] The acquisition step of the instantaneous morphological similarity is:

[0061] According to the micro-morphological feature vector set, slide one bit in time sequence to extract two adjacent feature vectors, align them one by one according to the component index, fill in the missing components by interpolation with adjacent effective components, record the pair order and time identifier, and obtain a sequence of adjacent feature vector pairs.

[0062] According to the sequence of adjacent feature vector pairs, establish a monotonic matching path of components to components for each pair, calculate the Fréchet distance value of each pair, normalize the Fréchet distance value to a similarity value by linear mapping from zero to one, and obtain the instantaneous morphological similarity.

[0063] Specifically, according to the micro-morphology feature vector set, the set is arranged in time sequence, where each row is a vector representing the features of a wheel passing event. During processing, a sliding window with a size of 2 is used, starting from the beginning of the data set and moving one vector position back each time. Within each window, the two adjacent micro-morphology feature vectors in time are extracted, for example, the first and second vectors are extracted the first time, the second and third vectors are extracted the second time, and so on. For each pair of extracted vectors, alignment is required according to the internal component index, i.e. the first component value is paired with the first component value, the second is paired with the second, and so on until the last component. In actual operation, due to signal interference or calculation anomalies, some components in a feature vector may fail to be successfully calculated and appear as missing values, for example, a vector may show as [0.85, 15.2, missing, 0.3, 40.1, 0]. When this happens, linear interpolation is used for filling. Specifically, the nearest two valid values before and after the missing component in time are found, for example, if the third component of the nth vector is missing, then the corresponding third components of the (n-1)th and (n+1)th vectors are found, for example, -14.5 and -15.5 respectively, then the average of these two values, -15.0, is used as the interpolated value. If there is only one valid value before and after, then the valid value is directly used for filling. After alignment and interpolation, the pair of vectors and their common time identifier, for example, the timestamp of the later vector, are recorded as a data unit. The sliding window is continuously moved and the above operations are repeated until the entire micro-morphology feature vector set is traversed, obtaining the sequence of adjacent feature vector pairs.

[0064] According to the adjacent feature vectors, each pair of feature vectors in the sequence is processed, each pair of vectors, such as vector P and vector Q, is regarded as two discrete sequences composed of its components, the Fréchet distance between them is calculated, and the calculation process is realized by constructing a cost matrix, the rows and columns of the matrix correspond to the components of vector P and vector Q respectively, and the value of each cell in the matrix is the square of the Euclidean distance between the corresponding component values. Then, a dynamic programming algorithm is used to find a path that can only move to the right, up or right up from the lower left corner of the matrix to the upper right corner, so that the maximum value in all cells on the path is minimized. The minimum value found is the Fréchet distance value between the two feature vectors. After obtaining the Fréchet distance value, it needs to be converted into a more intuitive similarity index, and this process is completed by linear mapping. First, the maximum and minimum distance boundaries need to be determined, and these two boundary values are obtained by calibrating a large amount of historical data collected under the normal operation state of the device. Specifically, all adjacent feature vector pairs in continuous operation for one month are collected, the respective Fréchet distances are calculated, and a distance distribution is formed. The 0.5 percentile point of the distribution is taken as the minimum distance boundary, for example 0.05, and the 99.5 percentile point is taken as the maximum distance boundary, for example 2.50. Then, for each newly calculated Fréchet distance value, a linear mapping formula is applied for conversion. The similarity value is calculated as 1 minus the quotient of (current Fréchet distance value minus minimum distance boundary) and (maximum distance boundary minus minimum distance boundary). For example, if the currently calculated Fréchet distance is 0.30, its similarity value is 1 - (0.30 - 0.05) / (2.50 - 0.05) = 0.898. All calculated similarity values are normalized to the interval of 0 to 1, where 1 represents complete identity and 0 represents maximum difference. Finally, the instantaneous shape similarity is obtained.

[0065] The acquisition step of the reflection consistency score is:

[0066] According to the instantaneous shape similarity, the reflection consistency score is calculated in combination with the total score of the previous period, and the calculation formula is:

[0067] , ;

[0068] Among them, is the reflection consistency score, is the total score of the previous period, is the instantaneous shape similarity, is the dynamic historical weight, is the basic score sensitivity coefficient, is the change impact sensitivity coefficient, is the score degradation degree, is the instantaneous change impact.

[0069] Specifically, in the reflection consistency score calculation formula, the first formula is the update equation of the score, which performs a weighted average of the historical score and the current instantaneous similarity , and the weight is dynamically changed and determined by the second formula, which makes the weight very sensitive to changes in two key factors, the first factor is the score degradation degree , that is, the difference between the historical score and the full score 1, the lower the historical score, the greater this item, resulting in a smaller weight , thereby more new similarity values are adopted ; the second factor is the instantaneous change impact , that is, the difference between the current similarity and the historical score, the greater the difference, the more severe the state change, and the smaller the weight , which also increases the influence of new data. This design enables the system to adaptively adjust the dependence on historical information and current information, achieving smooth tracking of slowly changing faults and capturing of sudden abnormalities.

[0070] is the total score of the last period, which is the reflection consistency score obtained at the end of the last calculation period, as an iterative value, reflecting the cumulative health status of the device up to the last time, its value range is between 0 and 1, at the first start or reset, an initial value is set to 1.0, representing that the device is completely healthy in the initial state, and in each subsequent period, the value is directly taken from the calculated in the previous step, without additional collection, for example, if the reflection consistency score calculated in the t-1 period is 0.95, then in the calculation of the score in the t period, the value of is 0.95;

[0071] is the instantaneous similarity, which is obtained by calculating the Fréchet distance between the micro-morphology feature vectors of the two consecutive wheel signals obtained at the latest, and then linearly mapping it to the interval [0, 1], directly quantifying the similarity between the latest wheel passing event and the immediately preceding event in signal morphology, which is a core indicator reflecting the current instantaneous working state of the device, the value is calculated by the previous step, for example, after calculating the Fréchet distance of the latest adjacent feature vector pair sequence and linear mapping, an instantaneous similarity value of 0.80 is obtained, then the value of in this calculation is 0.80;

[0072] is a hyper-parameter used to adjust the degree of influence of historical scores on the weight, which determines the sensitivity of the system to long-term and slow performance degradation. The value of is set based on analysis and tuning of historical failure data. Specifically, a dataset containing known slow-degradation failure cases (e.g., slow decline of sensor sensitivity over time) is collected, in which the device running score should slowly decline before failure occurs. Through repeated experiments, the value of is adjusted and the scoring algorithm is run, and the shape of the score decline curve is observed. The goal is to find a value of that makes the score curve show a clear and monotonous downward trend within a reasonable warning period before failure occurs. For example, through simulation testing, it is found that when is set to 2.0, for typical slow-degradation failures, the warning system can stably trigger the observation state 5 days before failure occurs. Therefore, the final value of is determined to be 2.0.

[0073] is a parameter used to adjust the degree of influence of instantaneous changes on the weight, which determines the response speed of the system to sudden events or dramatic changes in signal pattern. The setting process of is similar to that of, but a dataset containing sudden failure cases (e.g., physical impact on the sensor causing signal distortion) is used. In these cases, the signal similarity will drop sharply at the time of failure. By adjusting the value of, the goal is to make the reflection consistency score also have a significant jump within the period of sharp similarity drop, so that it can be quickly identified. For example, testing on data containing sudden impact events, it is found that when is set to 5.0, an event that causes the instantaneous pattern similarity to drop from 0.95 to 0.70 can make the reflection consistency score drop by more than 10% within a single period, meeting the requirement of fast response. Therefore, the final value of is determined to be 5.0.

[0074] According to the parameters, the calculation is as follows:

[0075] Set a calculation period t, and obtain the values of the parameters from the system:

[0076] Running total score of the previous period ;

[0077] Instantaneous pattern similarity of the current period ;

[0078] Basic score sensitivity coefficient ;

[0079] ​​​​​​​​Coefficient of change impact sensitivity ;

[0080] First, the dynamic historical weight is calculated :

[0081] ;

[0082] :

[0083] :

[0084] :

[0085] :

[0086] :

[0087] Then, the current reflection consistency score is calculated :

[0088] :

[0089] :

[0090] :

[0091] :

[0092] :

[0093] The result shows that the reflection consistency score of the shafting equipment at the current period t is 0.86411, which is the evaluation result after integrating the long-term health status of the equipment (the total score of the previous period is 0.95) and the current instantaneous working performance (the instantaneous shape similarity is 0.80). The score decreases from 0.95 to 0.86411, indicating that the state of the equipment has deteriorated significantly. This decrease is mainly caused by the sudden drop of the instantaneous shape similarity from a high level (implied in the historical score of 0.95) to 0.80. The value of 0.86411 will be used as the of the next calculation period and will be used for subsequent trend analysis and early warning judgment. If the score continues to decline, it may trigger a failure warning.

[0094] The acquisition step of the score deterioration trend determination result is:

[0095] According to the reflection consistency score, the reflection consistency scores of multiple consecutive periods are summarized in chronological order, the missing values are removed and the boundary vacancies are linearly interpolated to form a continuous period reflection consistency score sequence.

[0096] Based on the continuous periodic reflectivity consistency score sequence, the weighted trend coefficient is calculated using the following formula:

[0097] ;

[0098] in, This is a weighted trend coefficient. For the number of cycles, For the first Periodic index, For the first Periodic index, For the first The reflection consistency score for each cycle, For the first The reflection consistency score for each cycle, The time decay factor, As an amplitude-sensitive factor, For the first The interval between each cycle and the end;

[0099] Based on the weighted trend coefficient, a degradation trend threshold is set and compared with the weighted trend coefficient. If the weighted trend coefficient is less than the degradation trend threshold, degradation is marked; otherwise, non-degradation is marked, and a degradation trend judgment result is generated.

[0100] Specifically, based on the reflection consistency score, the length of the period window used for trend analysis is first determined. This length is set based on the typical failure evolution time. For example, according to historical maintenance records, a slow-change failure of axle counting equipment typically takes about 30 days from the appearance of initial symptoms to complete failure. Therefore, the number of consecutive periods is set to 30, that is, the reflection consistency scores of the past 30 calculation periods are summarized to form an initial time series. During the summarization process, some period score data may be missing due to communication interruptions or other reasons. For these internal missing values, linear interpolation is used to fill them in. For example, if the score of the 15th period is missing... Given that the score for the 14th period is 0.92 and the score for the 16th period is 0.90, the score for the 15th period is interpolated to be the average of these two values, which is 0.91. In addition to internal missing values, there may also be boundary gaps, that is, the number of data at the beginning or end of the sequence is less than 30, such as after the system has just started or restarted. In this case, forward or backward padding is used to fill the gaps. Specifically, if there is a lack of data at the beginning of the sequence, the first valid score value in the sequence is used to fill all the gaps in front. After processing all missing values, a complete data sequence of length 30 with no missing values ​​is obtained, forming a continuous periodic reflection consistency score sequence.

[0101] In the formula of the weighted trend coefficient, the time decay weight is introduced to make the recent numerical change have greater influence on the trend judgment than the long-term change. The double summation is used to traverse all the scores in sequence in time , and the difference value is calculated. The sign of the difference value directly reflects the increasing or decreasing trend of the local interval. The hyperbolic tangent function tanh is used to process the difference value, which can compress the difference value of any size to the interval of -1 to 1, thereby playing a normalization role, so that the trend judgment is not excessively affected by individual violent fluctuations, and more attention is paid to the consistency and persistence of the change. Meanwhile, the exponential decay term is introduced as the weight, where represents the distance of the data point from the end of the sequence. The closer the distance, the greater the value of the term, and the greater the weight. This ensures that the recent device state change has a higher proportion in the trend evaluation. Finally, all the weighted trend items are summed and divided by the total weight to obtain a standardized weighted trend coefficient value, which can sensitively reflect whether there is a persistent, especially recent, deterioration trend in the device health score.

[0102] is the number of periods, which defines the size of the time window for trend analysis, i.e., the length of the sequence of consecutive period reflection consistency scores. The selection of the value needs to balance the sensitivity and stability of trend detection. A smaller value, such as 7, can quickly respond to recent changes but is easily disturbed by short-term noise. A larger value, such as 30, can smooth the noise and more reliably identify long-term trends, but it responds slowly to new changes. The value of here is set to 30 according to the setting of the previous step, which matches the regular monthly inspection period of railway equipment. In the following example, a smaller window is selected.

[0103] and are the reflection consistency scores of the th and th periods, respectively. These two parameters are the basic elements that constitute the sequence of consecutive period reflection consistency scores, which are calculated and summarized by the previous step. They are direct quantitative indicators for evaluating the health status of the device at a specific time point, with a value range of 0 to 1. In this calculation, the sequence of consecutive period reflection consistency scores generated in the previous step is used. For example, when , a specific sequence is , where , , , ​;

[0104] is a time decay factor, which is a hyper-parameter that determines how much weight is given to recent data, the larger the value, the faster the time weight decays, and the more recent data the trend calculation is biased towards, this parameter is determined by grid search optimization on a historical dataset containing known slowly developing faults, by testing a series of values (e.g. from 0.01 to 1.0) and evaluating the early warning performance of the calculated weighted trend coefficients before the faults occurred, the value that gives the earliest and most stable negative trend indication is chosen . After extensive back analysis on historical data, it is found that ;

[0105] is a magnitude sensitivity factor, this parameter is used to adjust the sensitivity of the hyperbolic tangent function tanh to the score difference, the larger the value, the steeper the tanh function curve near zero, making even a small score difference amplified to near -1 or 1, thus making the trend calculation closer to a sign test that only cares about direction but not magnitude, on the contrary, a smaller value preserves more information about the score difference magnitude, the setting of this parameter is also based on historical data tuning, by analyzing the range of small fluctuations in normal operation scores and the decline amplitude of real early degradation scores, a value that can effectively distinguish the two is chosen, for example, normal fluctuations are usually less than 0.01, while early degradation decline amplitude is around 0.02, set so that tanh(100*0.01) has a small value, while tanh(100*0.02) has a value close to 1, thus effectively amplifying the degradation signal, therefore, set ;

[0106] According to the parameters, the calculation is as follows:

[0107] Given parameter values:

[0108] Number of periods .

[0109] Sequence of consecutive period reflection consistency scores .

[0110] Time decay factor .

[0111] Magnitude sensitivity factor .

[0112] First, calculate the denominator, which is the total weight:​

[0113] ;

[0114] When i = 1: , the weight is .

[0115] When i = 2: , the weight is .

[0116] When i = 3: , the weight is .

[0117] Total weight = .

[0118] Then the sum of the weighted trend items, i.e. the molecule, is calculated.

[0119] Finally, the weighted trend coefficient is calculated:

[0120] ;

[0121] The result shows that, in the selected 4 cycles, the reflection consistency score presents a relatively obvious downward trend, and the weighted trend coefficient is -0.7030. The negative value indicates that the overall trend is downward, and the absolute value is relatively large, indicating that the downward trend is consistent and continuous. If the value is lower than the preset degradation trend threshold, it can be considered that the device has the risk of slow-changing failure.

[0122] According to the weighted trend coefficient, a degradation trend threshold needs to be set for comparison, and the setting process of the threshold is based on statistical analysis of a large amount of historical operation data. Specifically, two types of data sets are collected, one is the sequence of weighted trend coefficients calculated by the equipment in the long-term healthy running state, and the other is the sequence of weighted trend coefficients within 30 cycles before the confirmed slowly varying failure. The coefficient sequence in the healthy state is statistically analyzed, and its distribution is usually centered on 0 and presents an approximate normal distribution. The mean and standard deviation are calculated, for example, the mean is 0.05 and the standard deviation is 0.15. The coefficient sequence in the failure precursor period is statistically analyzed, and its distribution will deviate to the negative value area, for example, the mean is -0.6 and the standard deviation is 0.2. The setting goal of the degradation trend threshold is to maximize the identification of true degradation trend while minimizing the false positives caused by normal data fluctuations. According to the 3-sigma principle, the threshold is set outside the negative boundary of the healthy data distribution, for example, set to the healthy state mean minus three times the standard deviation, that is, 0.05 - 3 * 0.15 = -0.40. This calculated -0.40 is used as the degradation trend threshold. Then, the latest calculated weighted trend coefficient is compared with the threshold. If the weighted trend coefficient is less than -0.40, it is determined that there is a degradation trend, and the state is marked as “degradation”. Otherwise, if the weighted trend coefficient is greater than or equal to -0.40, it is determined as “non-degradation”. Finally, the score degradation trend determination result is generated.

[0123] The acquisition step of the slowly varying failure warning state is:

[0124] According to the reflection consistency score, the latest reflection consistency score is read, the warning threshold is called for single comparison, the relationship is marked as less than, equal to or greater than, and the difference between the latest reflection consistency score and the warning threshold is calculated. The comparison timestamp and relationship label are recorded to obtain the reflection consistency score threshold comparison result.

[0125] According to the reflection consistency score threshold comparison result, the score degradation trend determination result is read. If the relationship label is less than and the score degradation trend determination result is yes, it is set as triggered. If the relationship label is less than and the score degradation trend determination result is no, it is set as observed. If the relationship label is equal to or greater than, it is set as not triggered. The warning trigger decision result is generated.

[0126] According to the warning trigger decision result, the trigger is mapped to the slowly varying failure warning state of the counting shaft equipment as warning, the observation is mapped to the slowly varying failure warning state of the counting shaft equipment as observation, and the non-trigger is mapped to the slowly varying failure warning state of the counting shaft equipment as normal. The slowly varying failure warning state of the counting shaft equipment is output.

[0127] Specifically, according to the reflection consistency score, first, the latest calculated reflection consistency score value is extracted from the time series data storage, for example, the latest score obtained is 0.86411, then a preset warning threshold is called, the setting of the warning threshold is based on risk assessment and operation requirements, specifically, by analyzing the distribution of the reflection consistency score when the confirmed fault occurs in the historical data, and combining the definition of the equipment state level in the maintenance procedure, for example, the score below 0.90 is defined as the state that needs attention, below 0.85 must be checked, select a value between the two with a certain safety margin as the warning threshold, for example, 0.88, this threshold aims to give an early warning before the equipment state enters the dangerous area that needs to be checked, compare the latest score 0.86411 with the warning threshold 0.88, because 0.86411 is less than 0.88, so the relationship label is marked as "less than", at the same time, calculate the difference between the two, that is, 0.86411 minus 0.88 equals -0.01589, this difference can quantify the degree of score deviation from the threshold, finally, record the system timestamp of this comparison, for example, "2023-10-27 15:30:00", and the relationship label "less than" just generated together, form a complete comparison record, get the reflection consistency score threshold comparison result.

[0128] According to the reflection consistency score threshold comparison result, first, the latest relationship label is read from the result, for example, the label is "less than", then, the latest score degradation trend judgment result is obtained from the trend analysis module of the previous step, for example, the judgment result is "degradation", then, according to the preset decision logic, the logic rule is as follows, rule one, if the relationship label is "less than" and the score degradation trend judgment result is "degradation", that is, the latest score has fallen below the warning threshold, and the score has shown a continuous downward trend in the past period of time, then the decision result is set to "trigger", rule two, if the relationship label is "less than" but the score degradation trend judgment result is "non-degradation", that is, although the latest score is below the threshold, this decline may be caused by a single sudden disturbance, and there is no continuous degradation trend, then the decision result is set to "observe", rule three, if the relationship label is "equal to" or "greater than", that is, the latest score is still within the safe range, then no matter the trend, the decision result is set to "not trigger", in this example, the relationship label is "less than" and the trend judgment is "degradation", which meets the conditions of rule one, therefore, the warning trigger decision result is set to "trigger".

[0129] According to the early warning trigger decision result, a state mapping operation is performed, which defines the correspondence between the decision result and the final output device early warning state. The mapping relationship is fixed, specifically, the decision result "trigger" is mapped to the "early warning" level of the gradual failure early warning state of the axle counter device, the decision result "observe" is mapped to the "observe" level, and the decision result "not trigger" is mapped to the "normal" level. According to the decision result "trigger" obtained in the last step, the mapping relationship is consulted to determine that the gradual failure early warning state of the current axle counter device is "early warning". This "early warning" state information will be sent to the central monitoring system of the railway to prompt the maintenance personnel in the form of highlighting or sound alarm. At the same time, detailed diagnostic information, including the reflection consistency score triggering the early warning, the score degradation trend determination result and the related historical data curve, will also be packaged and recorded for the maintenance personnel to conduct in-depth analysis and formulate maintenance plans. Finally, the determined "early warning" state is output as the final result of the current diagnosis process.

[0130] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any skilled person in the art can modify or change the above disclosed technical content to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments without departing from the technical solution content of the present application still belongs to the protection scope of the present application.

Claims

1. A method for online fault diagnosis of a railway axle counter equipment, characterized in that, The method comprises the following steps: Online real-time acquisition of the original analog voltage signal of the axle counting sensor, sampling of each wheel passing signal to obtain original voltage signal samples, time axis stretching transformation of each original voltage signal sample, and establishment of a standardized waveform sequence; According to the standardized waveform sequence, parameters of each standardized waveform in the sequence are extracted to generate a multi-dimensional feature parameter group, and each multi-dimensional feature parameter group associated with the state of the sensor coil is combined into a micro-morphology feature vector set; According to the micro-morphology feature vector set, two continuous feature vectors are sequentially taken out, the Fréchet distance between the two feature vectors is calculated, the instantaneous morphological similarity is obtained, the instantaneous morphological similarity is operated with the running total of the last period, and a reflection consistency score is obtained; According to the reflection consistency score, the numerical change in a plurality of continuous periods is tracked, it is determined whether the numerical value presents a monotonic decreasing trend, a score degradation trend determination result is generated, the latest reflection consistency score is compared with a preset warning threshold, the score degradation trend determination result is combined, and a slow-changing fault warning state of the axle counting equipment is output; The score degradation trend determination result is obtained by: According to the reflection consistency score, the reflection consistency scores of a plurality of continuous periods are sequentially collected, missing values are removed and boundary gaps are linearly interpolated to form a continuous period reflection consistency score sequence; According to the continuous period reflection consistency score sequence, a weighted trend coefficient is calculated, and the calculation formula is: ； wherein, is a weighted trend coefficient, is a number of cycles, is a first cycle index, is a first cycle index, is a first cycle reflection consistency score, is a first cycle reflection consistency score, is a time decay factor, is an amplitude sensitivity factor, is a first interval from end of cycle; According to the weighted trend coefficient, a degradation trend threshold is set and compared with the weighted trend coefficient, if the weighted trend coefficient is less than the degradation trend threshold, it is marked as degradation, otherwise it is marked as non-degradation, and a score degradation trend determination result is generated; The slow-changing fault warning state is obtained by: According to the reflection consistency score, the latest reflection consistency score is read, the warning threshold is called for single comparison, the relationship is marked as less than, equal to or greater than, the difference between the latest reflection consistency score and the warning threshold is calculated, the comparison time stamp and the relationship label are recorded, and a reflection consistency score threshold comparison result is obtained; According to the reflection consistency score threshold comparison result, the score degradation trend determination result is read, if the relationship label is less than and the score degradation trend determination result is yes, it is set as triggered, if the relationship label is less than and the score degradation trend determination result is no, it is set as observed, if the relationship label is equal to or greater than, it is set as not triggered, and a warning trigger decision result is generated; According to the warning trigger decision result, the trigger is mapped to the slow-changing fault warning state of the axle counting equipment as warning, the observation is mapped to the slow-changing fault warning state of the axle counting equipment as observation, and the non-trigger is mapped to the slow-changing fault warning state of the axle counting equipment as normal, and the slow-changing fault warning state of the axle counting equipment is output.

2. The on-line fault diagnosis method for railway axle counter equipment according to claim 1, characterized in that, The standardized waveform sequence is obtained by: According to the original analog voltage signal of the axle counting sensor, the start and end time indexes of each wheel passing signal are located, the continuous voltage value segment is cut off according to the start and end time indexes, the voltage value is recorded according to the fixed sampling interval, and the original voltage signal sample is obtained; According to the original voltage signal sample, the proportion of each time span to the target time length is calculated, the time axis is stretched or compressed in proportion, the peak position sequence and zero position are kept consistent in uniform time step resampling, and a voltage value sequence after time axis stretching is generated; According to the voltage value sequence after time axis stretching, the same length sample from the same starting point to the same ending point is intercepted, the sequence order is arranged according to the same sample number, and the starting zero value is aligned to form a standardized waveform sequence.

3. The on-line fault diagnosis method for railway axle counter equipment according to claim 1, characterized in that, The acquisition step of the multi-dimensional feature parameter group is: According to the standardized waveform sequence, the amplitude of each standardized waveform is read, the peak sample position is located, the zero-crossing point position is located, the starting point of the rising edge and the termination point of the falling edge are labeled at the same time, and the waveform structure labeling result is obtained; According to the waveform structure labeling result, the amplitude difference is counted in the left and right intervals of the peak sample position respectively to form the symmetry of the waveform peak, the amplitude increment and time span from the starting point of the rising edge to the peak sample position are formed to form the slope of the rising edge, the amplitude decrement and time span from the peak sample position to the termination point of the falling edge are formed to form the slope of the falling edge, and the oscillation decay rate is formed by comparing the ratio of adjacent amplitudes in the interval divided by a fixed step after the zero-crossing point. The waveform integral area is formed by point-by-point accumulation of all amplitudes under the uniform time step, and a multi-dimensional feature parameter group is generated.

4. The on-line fault diagnosis method for railway axle counter equipment according to claim 1, characterized in that, The acquisition step of the micro-morphology feature vector set is: According to the multi-dimensional feature parameter group, the sensor coil state identifier corresponding to each standardized waveform is read, the sensor coil state identifier is attached to the end of the multi-dimensional feature parameter group, and a micro-morphology feature vector set is formed.

5. The on-line fault diagnosis method for railway axle counter equipment according to claim 1, characterized in that, The acquisition step of the instantaneous morphology similarity is: According to the micro-morphology feature vector set, the adjacent two feature vectors are extracted by sliding one bit in time sequence, aligned one by one according to component index, the missing components are filled by interpolation with adjacent effective components, and the order and time identifier are recorded. Get the sequence of adjacent feature vector pairs; According to the sequence of adjacent feature vector pairs, a monotonic matching path of components to components is established for each pair, the Fréchet distance value of each pair is calculated, the Fréchet distance value is normalized to a similarity value by linear mapping from zero to one interval, and the instantaneous morphology similarity is obtained.

6. The on-line fault diagnostic method for railway axle counter equipment according to claim 1, characterized in that, The acquisition step of the reflection consistency score is: According to the instantaneous morphology similarity, the reflection consistency score is calculated combined with the total score of the last period.

7. The on-line fault diagnosis system of the on-line fault diagnosis method of a railway axle counter device according to any one of claims 1 to 6, characterized in that, It includes: The acquisition module is used for collecting the original analog voltage signal of the axle sensor online in real time, sampling the signal of each wheel passing, obtaining the original voltage signal sample, and performing time axis stretching transformation on each original voltage signal sample to establish a standardized waveform sequence; The feature extraction module is used for extracting parameters from each standardized waveform in the standardized waveform sequence to generate a multi-dimensional feature parameter group, and combining each multi-dimensional feature parameter group associated with the sensor coil state into a micro-morphology feature vector set; The similarity calculation module is configured to: according to the micro-morphology feature vector set, sequentially take out two continuous feature vectors, calculate the Fréchet distance between the two feature vectors, obtain an instantaneous morphological similarity, and perform operation on the instantaneous morphological similarity and a total score of a previous cycle to obtain a reflection consistency score; The early warning judgment module is configured to: track a numerical change in a plurality of continuous cycles according to the reflection consistency score, judge whether the numerical change presents a monotone decreasing trend, generate a score degradation trend judgment result, compare the latest reflection consistency score with a preset early warning threshold, combine the score degradation trend judgment result, and output a slow change fault early warning state of the axle equipment.

Images