AT522937B1 - Procedure for detecting technical irregularities in railway vehicles from the analysis of sound and vibration data - Google Patents

Abstract

The invention relates to a computer-implemented method for detecting technical irregularities on a rail vehicle, - the noise emissions generated by the rail vehicle while passing the measurement cross-section are detected, - the measurement period encompassing the time period in which the rail vehicle completely passes the measurement cross-section, - wherein a three-dimensional data structure (fsi3(t, fband, frep)) is present as the result of a mathematical analysis step, - whereby in a decision step based on the three-dimensional data structure (fsi3(t, fband, frep)) a from the respective frequency band (fband), the time (t) and the repetition frequency (frep). dependent flat area indicator (fsi) is determined, - the flat area indicator (fsi) indicating the time ranges associated with technical irregularities within the measurement period (tMess), and - the flat area indicator (fsi) being determined by - the individual frequency bands (fband) being combined , so that a two-dimensional data structure (fsi2(t, frep)) is present.

Description

description

PROCEDURE FOR DETECTING IRREGULARITIES

The invention relates to a computer-implemented method for detecting technical irregularities in a rail vehicle, in particular due to flat spots on the running surface of a wheel of the rail vehicle and a method for detecting technical irregularities in railway vehicles from the analysis of sound and vibration data.

According to the invention, a method is used to analyze sound or vibration waves emitted by one or more rail vehicles (“train”) while driving with a view to detecting possible technical irregularities on one wheel or more wheels of the vehicles, which thereby pose a potential risk to the wheel and rail and axle and/or rail breakage may occur.

[0003] Such irregularities are often caused by so-called “flat spots” on the running surface of the wheel of a rail vehicle. These are usually caused by a temporary blockage of the brakes on a wheel during braking or by stones thrown up. This causes the wheel to drag on the rail and can wear away part of the tread, causing the wheel to lose its ideal circular circumference. After the blockage has been lifted, this flat spot(s) hit the rail with each revolution of the wheel during the further journey(s) and produce a characteristic hitting noise ("knocking" / "hammering"). Such situations are recognized according to the state of the art:

(a) By visually inspecting the stationary train ("car master inspection"),

(b) By "listening" to the train while it is passing by an employee of the railway company (e.g. station master in train stations) or

(c) By measuring the wheel geometry with complex and expensive trackside measuring systems (so-called "train monitoring systems") in the route network.

[0004] This is cumbersome, not very accurate, cost-intensive, and not very efficient if only because of the low “density” of the test points in the railway network.

Similar methods or methods known from the prior art are described, for example, in CN 101697175 A, EP 2393026 A2, CN 102880767 A, AT 511769 A2, CN 107292046 A or CN 107215353 A.

The method according to the invention is based on a sequence of automatically recorded and digitized acoustic emission values which the rail vehicle(s) emits while passing the measuring point.

According to the invention, the following method is provided:

[0008] Computer-implemented method for detecting technical irregularities on a rail vehicle, in particular caused by flat spots on the running surface of a wheel of the rail vehicle, - using a sensor positioned on a measuring cross-section arranged at right angles to the track axis of the route of the rail vehicle, during a measuring period (tmess) the noise emissions generated by the rail vehicle while passing the measurement cross-section are detected, - the measurement period (tmess) encompassing the time range in which the rail vehicle completely passes the measurement cross-section, - wherein in an analysis step - the detected in the measurement period (tmess). signal is divided into a number of frequency bands (fband), - the chronological derivation of the course of the signal energy of the signal in the frequency bands (foana) is determined, - the determined chronological derivation of the course of the signal energy in the individual frequency

bands (fvana) Each divided into a number of time intervals and an autocorrelation

is subjected to

- the autocorrelated signal of the individual frequency bands (fvana) and the individual time intervals

valle, in particular by means of a Fast Fourier Transform, on the existence

of regularly repeating components is examined, - the result of the mathematical analysis step being a three-dimensional data structure (fSi3(t, foand, frep)) whose indices are defined by

- an index characterizing the measurement time (t).

- an index identifying the frequency band (foana) and

- one the repetition frequency (fıep) per frequency band and per measurement time (Fpiand)

distinctive index, and

- wherein in a decision step based on the three-dimensional data structure (fsi3(t, foand, frep)) one of the respective frequency band (feana), the time (t) and the repetition frequency (fee). dependent flat spot indicator (fsi) is determined,

- where the flat spot indicator (fsi) is the time frame associated with technical irregularities

ranges within the measurement period (tmess), and

- where the flat indicator (fsi) is determined by - merging the individual frequency bands (foana) so that a two-dimensional data structure (fsi2(t, frep)) is present, the indices of which are defined by

- an index characterizing the measurement time (t) and

- an index characterizing the repetition frequency (frepe), - repetitions are identified by recording the temporal dimension in the two-dimensional data structure (fsi2(t, frep)) and analyzing each point in time as a vector fsi(freg) as a function of the repetition frequency (fıep). is determined as part of the analysis, the maximum of the vector fsi(frep) is determined and used as the basis for a threshold value, and neighboring values that exceed the threshold value are combined into groups and then analyzed using statistical methods, so that finally a number of There are result vectors, each dependent on time and extending over the measurement period (tmess), the result vectors comprising:

- the largest number of consecutive elements in a group,

- the maximum relative deviation of the repetition frequencies (fıep) of all groups that

are not the first, at an integer multiple of the repetition frequency of the first

Group,

- the repetition frequency (fep) of the first group, and - a presence of a technical irregularity is identified if

- the maximum relative deviation of the repetition frequencies one for the maximum re-

lative deviation falls below the specified threshold value,

- the largest number of consecutive elements in a group one for the element

number falls below the predetermined threshold and

- Repetition frequency (fiepe) for the repetition frequency (fee) specified

falls below the threshold.

[0009] As a result, one or more indicators are delivered within the period under consideration, the value(s) of which serves as an indication of the existence of the irregularities sought and can be used for further interpretation (see “Flat Spot Indicator”/FSI). The process described minimizes the effort involved in setting up and operating the measuring devices and brings about a qualitative and quantitative improvement in monitoring. This means that it is possible to react in good time to possible negative side effects, e.g. axle/rail breakage, vehicle failure, etc., and the process contributes to safe, efficient and highly available (rail) operation.

OBTAINING THE OUTPUT DATA:

The derivation of the basic data is basically shown in FIG. 1 and is not part of the application.

The method according to the invention is based on the evaluation of the sound or vibration emissions ("train noise") emitted by the rail vehicles of a train while driving, which is primarily caused by the rolling noise caused by the wheel/rail up to speeds of 200 km/h. contact is generated.

The (sound) emission emitted by the railway vehicle(s) of a train (1) as it passes the measurement cross section Q (2) is detected by a stationary electroacoustic pickup (microphone) (3) positioned on the measurement cross section Q, which is perpendicular to the axis of the track. recorded and converted into a continuous electrical signal "A" (4a).

In principle, the following description assumes that the sound emission is recorded by an electroacoustic sensor. However, other signal sources representing the sound emission, such as acceleration sensors, fiber-optic pickups, strain gauges, and other sensors, can also serve as signal sources.

The analog signal is subsequently converted by an analog/digital converter (4) with an adjustable but fixed sampling interval tavıası IN temporally equidistant punctual digital measured values (4a). In practice, the tapıası sampling interval is at least selected in such a way that the original analog signal can be correctly reconstructed up to frequencies of at least 20 kHz (Shannon theorem).

The signal is recorded at least over a time range tyess, which includes the passage of the train or the railway vehicle to be measured - represented by the first and last axle of the vehicle/train - at the measurement cross-section Q plus an adjustable pre- (tvor) and Run-on time (tnaen) overscored. This time range is referred to below as the "measuring time" (tmeas) (3a). Depending on the sampling interval taptast and the measurement time tmess, the sampling of the analog signal now results in a sequence of measured values equidistant in time with tapıası along the time axis.

This sequence of values “measured value(to)” to “measured value(t”)” (5) forms the basis for the method according to the invention.

ANALYTICAL BASIS:

The analytical basis of the method described for the analysis and calculation of the "flat point indicator" ("FSI") is based on the task of repetitive - i.e. repetitive - step functions of the amplitude in the individual frequency bands in the frequency mixture of the recorded (sound) to identify signals. In their entirety, these step functions form the described "knocking noise" of a flat spot. The signal level does not necessarily have to significantly exceed the general level of noise emissions.

It can be represented by various mathematical models.

One of them is the "Diracstoss" (see relevant literature). A Dirac burst represents a single very short and very large (sound) impulse, i.e. a sharp increase in amplitude and subsequent decay of the (sound) signal in the time domain. A Dirac impact has the frequency distribution 1, which means that all frequencies are equally represented in it. It follows that the 'thumping' sound of a flat spot should theoretically be detectable in all frequency bands. In this approach, therefore, the detected signal aspect is examined for the presence of Dirac surges.

[0020] In addition to the modeling approach using a “Dirac collision”, other approaches for modeling step functions can also be used. In the following, the analysis is followed by modeling through a Dirac collision.

Another characteristic is a repetitive repetition of the occurrence of Dirac shocks over time due to the rotation of the wheel while driving. This repetition frequency depends on the wheel circumference and the current speed. The method according to the invention takes this into account.

The frequency spectrum considered can - but does not have to - be limited in practice to the range from 100 Hz to 20 kHz. This means that the signals can be recorded with common technical means - such as a weatherproof standard measuring microphone or corresponding sensors.

[0023] Likewise, the period of time tmeas under consideration can include the time it takes the entire train to pass by, or just a subset of the recorded and digitized measured values.

DESCRIPTION OF THE METHOD STEPS: The method is based on 2 method steps, the course of which is shown in FIG. 2: Method step 1 - "mathematical analysis":

A 3-dimensional result matrix is formed from the sequence of values measured value(to) measured value(tn) of the recorded and digitized (sound) signal in 4 consecutive steps. In the time domain, this contains information about the measurement time about possible repetitive step functions.

Method step 2 - "FSI decision maker":

This step adopts the result matrix from step 1 and uses defined criteria to decide whether and if yes, at what time a possible technical irregularity in the sound signal could be identified.

[0029] As a result, the method may identify time periods within the measurement time in which possible technical irregularities were detected - the "flat spot indicator" (fsi).

PROCEDURE STEP 1 - “MATHEMATICAL ANALYSIS”:

The 'mathematical analysis' performed in the first step is a series of different mathematical/technical analysis methods to identify a repetitive 'banging' noise.

For this purpose, the following steps are applied to the value sequence of the sound signal:

1) Division of the sampled signal curve into frequency bands 2) Time derivation

3) Autocorrelation of time intervals

4) Calculation of the repetition frequency

The result is a 3-dimensional vector fsi3(t,foana,frep) IN which t (x-axis) spans the measurement time, foana (y-axis) the frequency bands and f,ep (z-axis) the Has repetition frequencies per fovand and which can now contain an indication of possible repetitive jump functions. This vector is then forwarded to the second process step (“decision maker”) as an input.

PROCEDURE

1) Division into frequency bands

A first method is the division of the signal into frequency bands, which are divided logarithmically. The signal is subjected to a "Fast Fourier Transformation" (FFT). Then the powers of the FFT "bands" are added by grouping the frequency bands according to the logarithm.

A second method is the division into fractional octave bands. Bandpass filters with a fraction n of the bandwidth of an octave (factor 2/n of the cut-off frequencies) filter the audio signal in order to divide it into the different frequency bands, where the following formulas apply:

{1} fa(k) = foent * 27

(2) fu) = fo() + 226 {3} fi(k) = fe(k)/22u Deseeeennnnn Number of bands per octave

Kosaeeeee Index of a fractional octave band'

feet +... center frequency of the frequency band k=0

fe(K) ...... center frequency of frequency band k

fu(k) ..... Upper limit frequency of the frequency band k fi(k) ...... Lower limit frequency of the frequency band k fmax 0... Maximum occurring frequency

fmin + minimum occurring frequency

Kmax....... Index of the highest frequency band

km +44. Index of the lowest frequency band

The center frequency of a frequency band is given as a reference. Here it is usual to use the value 1000 Hz. Based on this, the center frequency, the upper and the lower limit frequency are calculated for all frequency bands.

Rearranging the formulas {1}-{3} gives the formulas {4} and {5} for the calculation of the maximum and minimum index k. The function 'ld' is logarithm dualis with base 2, 'floor' rounds down to the nearest integer and 'ceil' rounds up to the nearest integer.

{4} Kımax = floor (n * Id (Amex) — 3)

cent {5} Kmin = ceil (n * Id (Zein) + 3) The maximum and minimum frequencies fmax and fmin are parameters that must be specified for the calculation. fmax can be a maximum of half the sampling frequency of the microphone and fmin should not be selected lower than the lowest frequency that the microphone can still pick up. A good approach is fmin = 200 Hz and fmax = 20 kHz.

Procedure:

(1) Fixation of the number of bands per octave ("1") - can be freely selected. (2) First the minimum and maximum value of the index k (kmax and kmin) is calculated. (3) Every integer from kmin to kmax is an index k for a fractional octave band.

The following steps (4-8) are now carried out for each band:

(4) For each of these bands, the center frequency f.(k), the upper limit frequency f.(k) and the lower limit frequency fi(k) are calculated with the index k.

(5) The filter coefficients for a bandpass filter are now calculated using the upper and lower limit frequencies of a 'fractional octave band' (fu and fi) and the sampling frequency of the audio recording.

(6) Then the signal is filtered through a digital filter with the calculated filter coefficients.

(7) The filtered signal is now divided into sequence blocks with a length of n, where: n>0 and the form 2", ideally in the application n=256 samples, whereby the blocks overlap half, and in the time domain multiplied by a Hann window.

(8) Now the energy of the sound in each of these windows is calculated by squaring each sample value individually, adding all values in a window and then calculating the logarithm of them.

[0040] The result is a 2-dimensional matrix that supplies the time profile of the signal energy in different frequency bands.

{E1} [ S(frequency band, time) ]

2) Time derivative

In the next step, each frequency band is temporally derived. To detect a flat, the difference is taken over at least 8 samples

{E2} [D(i) = S(i) - S(i-8) ].

[0042] Since the characteristic of a flat spot is the rapid increase, only positive values are of importance. Negative results of the derivation are set to 0. This process is carried out individually for all frequency bands.

[0043] A 2-dimensional matrix is again supplied as the output, which supplies the derived time profile of the signal energy in different frequency bands.

{E3} [ D(frequency band, time) ]

3) Autocorrelation of time intervals

[0044] Autocorrelation is now used to examine whether peaks are repeated in the derived time signal. This step is carried out individually for each frequency band, which means that the signal to be processed is only dependent on time [ d(t) ].

The signal is divided into blocks of an adjustable length (e.g. approx. 1.5 seconds), the number of blocks should be a natural power of two (2, 4, 8, ...). With the parameters of the previous steps, blocks with e.g. 512 values are formed, which corresponds to about 1.56 seconds at a sampling frequency of 42 kHz. For the next block, the time is shifted by approx. 0.2 seconds, which in this example corresponds to a shift of 65 values.

Each block is then autocorrelated. Since the autocorrelation is symmetric, only shift values >= 0 are considered. This step creates a vector with the same number of values as before the autocorrelation.

[0047] This vector is now normalized by dividing each value of the autocorrelation by the value at the shift=0.

The result of this step is a 3-dimensional matrix. For each frequency band there is a vector every 0.2 seconds containing the results of autocorrelation as a function of lag.

{E4} A(frequency band, time, lag)

4) Calculation of the repetition frequency

It is then analyzed whether there is a regularly repeating component in the autocorrelated signal. This step is carried out individually for each frequency band and for each time segment. Therefore, the only input here is the vector of the autocorrelation, which starts with the shift equal to 0 and with the value 1 due to the normalization.

The FFT is now calculated from this vector in order to be able to consider regular signal parts. The first value of the autocorrelation is one. When the signal is shifted by the distance between two 'hammer hits', the autocorrelation is high again, with near zero in between. Because the waveform begins with a maximum, the signal you are looking for is a cosine.

After the FFT, the sine component in the time signal is represented by the imaginary part and the cosine component by the real part. Since only the cosine is of interest, only the real part of the result of the FFT - and only positive values - is used. All negative values are set to 0.

Since the input signal of the FFT is not complex (imaginary part=0), the output vector of the FFT is symmetrical. Therefore, only the lower half of the vector is considered further after the FFT, only positive frequencies, not negative ones.

This frequency vector is now normalized by dividing it by its length.

The result is a frequency vector. "Frequency" is the rate/second at which a tapping sound repeats, and the value represents the "significance" or "signal strength" for those repetitions.

RESULT OF MATHEMATICAL ANALYSIS

The mathematical analysis supplies a 3-dimensional matrix fsi3 as output. The last step was carried out individually for each frequency band and individually for each point in time within each frequency band. The result is again a vector which supplies values depending on the repetition frequency. This is how three dimensions (axes) come about.

{E5} fsi3 = [A(time, frequency band, repetition frequency) ]

[0056] Dimension time: Distance between the value groups on the time axis Example: If time blocks are formed with a shift of 0.2 seconds to one another in step 3, there are separate values on the time axis every 0.2 seconds.

Dimension repetition frequency: vector with repetition frequencies

At any point in time on the t-axis, there is a vector with different repetition rates for each frequency band. If a value in this vector is significantly high, it means that a "noise" repeats itself regularly at this point in time with the corresponding repetition rate.

Dimension frequency band

The complete analysis was carried out individually for each frequency band. When a high value occurs at a specific time with a specific repetition rate, the dimension of the frequency band corresponds to the frequency of the repeating ('banging') sound.

PROCESS STEP 2 - "FSI DECISION MAKER":

In the second method step, the result of the first method step is analyzed. The "flat spot indicator" fsi depends on the frequency band (foana), the time (t) and the repetition frequency (frep).

In a further 3 consecutive steps, the decision (indication) is made as to whether possible flat spots (number >=0) have been detected:

1. Elimination of dimension foana - "fpMerge"

The individual frequency bands are combined. A corresponding approach, based on the median from the statistics, is to take over the second-largest value of the dimension fovand TO. This ensures that, on the one hand, a single "outlier" in a frequency band is not noticeable and, on the other hand, that several low values do not "pull down" the mean value. This results in the 2-dimensional matrix fsi2(t, frep).

2, identification of repetitions - "frGet"

The time dimension is now recorded in the 2-dimensional matrix fsi2(t, fie) and each point in time is considered as a vector depending on the repetition frequency: fsi(frep).

The maximum is sought in this vector and taken as the basis for a threshold value. The threshold can vary according to the requirements. Neighboring values that exceed the threshold value are now combined into groups with 1-n elements. The repetition frequency of the element with the highest fsi value is now taken from each group. Now for the groups 2-m the factor of the respective repetition frequency is calculated for that of the group element "1", for this factor the relative deviation to the next whole number is calculated and the absolute value taken from it.

In this step, 5 one-dimensional result vectors were generated from the 2-dimensional input vector, which are only dependent on the time dimension and extend over the measurement time:

{E6} [frGet(Numgroups, following, deviation, fr, fsi ) ] Numgroups: Number of groups

following: the largest number of consecutive elements in a group deviation: the maximum relative deviation (without dimensions, value range 0.0 to 0.5) fr: the fundamental (= the f,ep of the 1st group)

fsi: The maximum flat spot indication value (dimensionless)

3. Indication of possible flat spots - "Detector"

The 3rd step is now the evaluation of the 5 vectors from step 2. The following criteria are used for this:

K1: maximum deviation not too large (e.g. deviation < 0.1) K2: "narrow frequency groups" (e.g. following < 4) K3: repetition frequency not too large (e.g. frep < 15)

{E7} [E6(K1+K2+K3) ]

If—depending on the selected parameters—criteria K1, K2 and K3 apply, then it can be a possible flat spot.

As a possible additional sharpening of the decision, other criteria such as

K4: minimum consecutive number of values that meet K1 to K3 and K5: a minimum threshold value that the maximum value fsimax must at least exceed

come into use. {E8} [fsi = E7/(K4+K5) ] = [fsij(tbeg,tend, indicatorvalue)]

EXAMPLE: An example of the invention is shown in FIG.

Figure 1 (top) shows the frequency spectrum of a train over the measurement period. F, are the individual frequency bands.

Figure 2 (middle) shows the result after step 2 (fbMerge) with repetitive step functions.

Figure 3 (below) the result of step 3 (frGet) for 2 of the result vectors fsi and frep.

After applying the criteria K1 to K5, the suspected area of one or more flat spot(s) on one wheel or more wheels in the train consists—represented by the line formed between the circular markings (4-6 time/s) (“ flat spot indicator"). In this example, the inspection even revealed 2 flat spots on one wheel each of two adjacent wagons.

By mapping these results to a train axle pattern recorded parallel to this over the measurement time and/or automatic car identification reader, it is possible to deduce the relevant car concerned and, if necessary, appropriate further measures can be initiated.

Claims (1)

patent claim 1. Computer-implemented method for detecting technical irregularities on a rail vehicle, in particular caused by flat spots on the running surface of a wheel of the rail vehicle, - whereby the noise emissions generated by the rail vehicle while passing the measurement cross-section are detected during a measurement period (tmess) by means of a sensor positioned on a measurement cross-section arranged at right angles to the track axis of the route of the rail vehicle, - the measurement period (tmess) defining the time range in which the rail vehicle passes the measurement cross-section completely, - whereby in an analysis step - the signal detected in the measurement period (tmess) is subdivided into a number of frequency bands (foana), - the time derivation of the progression of the signal energy of the signal in the frequency bands ( foana) is determined, - the determined time derivation of the course of the signal energy in the individual frequency bands (foana) is divided into a number of time intervals and subjected to an autocorrelation, - the autocorrelated signal of the individual frequency bands (fand) and the individual time interval e is examined in each case, in particular by means of a fast Fourier transformation, for the presence of regularly repeating components, - the result of the mathematical analysis step being a three-dimensional data structure (fsi3(t, foand, frep)) whose indices are defined by - an index characterizing the measurement time (t), - an index characterizing the frequency band (feana), and - an index characterizing the repetition frequency (free) per frequency band and per measurement time (Fbana), and - wherein in a decision step based on the three-dimensional data structure (fsi3( t, foand, frep)) a flat area indicator (fsi) dependent on the respective frequency band (frana), the time (t) and the repetition frequency (fıep) is determined, - whereby the flat area indicator (fsi) indicates the time ranges associated with technical irregularities within the measurement period (tmess) indicates, and - wherein the flat spot indicator (fsi) is determined by - the individual frequency bands (fvand) are brought together so that there is a two-dimensional data structure (fsi2(t, f,ep)) whose indices are defined by - an index identifying the time of measurement (t) and - an index identifying the repetition frequency (frepe). , - Repetitions are identified by recording the time dimension in the two-dimensional data structure (fsi2(t, fep)) and analyzing each point in time as a vector fsi(freg) as a function of the repetition frequency (f,ep), whereby in the context of the analysis the maximum of the vector fsi(f'ep) is determined and used as a basis for a threshold and neighboring values that exceed the threshold are grouped together and then analyzed using statistical methods, so that finally there is a number of result vectors, each are dependent on time and extend over the measurement period (tmess), with the result vectors n include: - the largest number of consecutive elements in a group, - the maximum relative deviation of the repetition frequencies (f,ep) of all groups that are not the first to an integer multiple of the repetition frequency of the first group, - the repetition frequency (fep) the first group, and - the existence of a technical irregularity is identified if - the maximum relative deviation of the repetition frequencies falls below a threshold value specified for the maximum relative deviation, - the largest number of consecutive elements in a group falls below a threshold value specified for the number of elements and - Repetition rate (frep) falls below a predetermined threshold for the repetition rate (fıep). 2 sheets of drawings