System and method for automatic detection of bearing defects

Abstract

A method for automatic bearing fault detection provides an algorithm for processing condition monitoring data comprising vibration harmonics of at least one bearing (105) coupled to a rotating shaft (109), the bearing (105) having an inner ring and an outer ring. The algorithm is used to confirm with a high degree of confidence whether a bearing fault is present or not. Figure 2

Description

Title of the invention: System and method for automatic detection of bearing defects State of the prior art

[0001] A train with all axles having the same axlebox design may be equipped with bearings from different manufacturers, which bearings, even if they have the same spatial envelope and the same "capacity", often have different frequencies of bearing defects due to slight changes in their internal geometry. Managing which bearing type to install on which axlebox with, for example, @ptitude Observer (@O) manufactured by Aktiebolaget SKF with headquarters in 41550 Gothenburg, Sweden and quickly updating a database is costly in terms of man-hours and prone to errors.

[0002] One type of system, manufactured by Aktiebolget SKF, headquartered in 41550 Gothenburg, Sweden, called "Insight Rail System", uses GPS linear speed and wheel diameter to determine shaft rpm values ​​for each measurement. However, during the lifetime of a train wheel set, the wheel profiles suffer from wear, dents, flat spots and fatigue squats (scrubs) and therefore require reprofiling as is or at regular intervals. Reprofiling is done by turning or grinding and reduces the wheel diameter each time (both wheels on the same axle are always kept at the same diameter) so that during the lifetime of an axle, its diameter is typically reduced by more than 10% compared to its new condition.Managing individual axle wheel diameters, before (on an Insight installation) and after each re-profile, and then quickly updating the @ptitude Observer (@O) database is costly in terms of man-hours and also error-prone.

[0003] Often an Insight Rail system, manufactured by Aktiebolaget SKF, is installed on a train whose bearings are halfway through their service life, so one or more bearings may already have some degree of exfoliation and the system must detect bearing exfoliations without requiring a learning period. Unlike @0 Protean systems manufactured by Aktiebolaget SKF, headquartered in 41550 Gothenburg, Sweden, which require both a learning period with a "healthy" bearing to establish trend characteristics and / or alarm thresholds. Summary of the invention

[0004] According to one or more non-limiting embodiments, a method of performing automatic bearing fault detection is provided. The method comprises receiving, by a processor from one or more sensors, condition monitoring data, the condition monitoring data comprising harmonics of vibrations of at least one bearing coupled to a rotating shaft, the bearing having an inner ring and an outer ring, the method further comprising;

[0005] a) receiving vibration data previously or sequentially transformed into the frequency domain and finally provided in the form of magnitude and frequency matrices, with an approximate shaft speed in revolutions / minute or revolutions / second,

[0006] b) Applying a peak determination method that determines individual peaks in the background noise, the actual amplitudes of the peaks, and their exact peak center frequencies, with the exact frequency and amplitude of each possible peak stored in a matrix.

[0007] c) Identifying all possible integer suborders of each peak up to a predefined number of orders, including the peak itself, that fall within a specific range of target fundamental frequency ranges, one for each defined type of defect, and storing these peaks with their harmonic rank and their theoretical fundamental frequency in the form of order matrices.

[0008] d) clustering the theoretical fundamental frequencies that fall within each specific target range, the frequencies being clustered by one or more of numerous methods, the theoretical fundamental frequencies having a predefined cluster size limit and the clusters having acceptable clustering characteristics by the number of peaks they have and / or the most stringent cluster sizes are organized into groups and all others as outliers, and the groups are stored in order matrices, and

[0009] e) identifying in the original list all peaks that could be possible sideband components of the peaks identified in step d), having a base delta frequency in a specific range of target sideband frequency ranges, one for each defined defect type, and up to + / - delta orders of a predefined sideband to determine all possible theoretical base delta sideband frequencies and storing these possible sideband peaks with the number of sidebands, the theoretical delta sideband frequency and the center peak ID index, which fall in each of the target sideband ranges as sideband matrices, and

[0010] f) applying a clustering method to their basic delta frequencies with a predefined cluster size limit and the clusters having acceptable clustering characteristics by the number of peaks they have and / or the strictest cluster sizes are organized into sideband groups and all others as outliers, and the sideband groups laterals are stored in matrices with references to their respective center frequency peak from the order matrices.

[0011] According to a second aspect, the method may undertake model clustering for each bearing inner or outer ring defect (or other types of defects). Three factors for model building are defined by 1) identification of harmonics as number of orders, 2) number of sidebands and 3) fundamental frequency range and frequency range for sidebands, wherein one fundamental component, none or few harmonic components, none or few sideband components having a delta frequency other than its fundamental frequency compose the possible clustered models.

[0012] According to a third aspect, the method may undertake a determination of state indicators, wherein the selection of the most plausible group may then be performed by (a) correlation with weighted model components, (b) presence or absence of components (c) RSS of group components (d) pattern recognition process on each group.

[0013] According to a fourth aspect, the method may undertake high confidence fault identification, wherein faults may be predicted with a high degree of confidence according to the "Most-out-of-N" (MooN) principle. That is, for each logic test, most of the results of a test set of N measurements N must be "positive", the MooN logic parameters being defined by a) experienced analysts, b) design of experiments (DOE) on known data, c) machine learning. An artificial intelligence algorithm to "learn" the parameters, in order to identify the faults that exist, is used.

[0014] According to another aspect, the method may undertake a conversion of the magnitude and frequency matrices into an acceleration enveloped by a data acquisition system, and / or a Hanning window is performed on the data before the FFT Fast Fourier Transform (FT), and / or a RFT or repetitive Fourier Transform (RFTep) known by the acronym "RFT repetitive Fourier Transform" is performed on data to generate spectral data, and / or the magnitude of the FFT data is calculated from the real and imaginary spectral components, "Magnitude" = SQRT(RealA2 + ImagA2), and / or the spectrum is truncated.

[0015] According to another aspect, the method may undertake a division of each peak frequency in step c) by a number from 1 up to the number of harmonics to be included in turn to determine all possible theoretical fundamental sub-frequencies

[0016] According to another aspect, the method may undertake a subtraction of each peak frequency identified in step b) by all other peak frequencies in turn and dividing the absolute value of the result by the number 1 up to the number of sidebands to be searched for to identify all possible sideband peaks.

[0017] In another aspect, the method may undertake cluster optimization by applying a method of removing from the cluster list clusters that are a sub-cluster of another cluster having a lower or higher fundamental frequency, such that, but not limited to, (a) the majority of the order peaks in the cluster are present as orders in another cluster of lower fundamental frequency, (b) the orders of the peaks are all multiples 2, 3 or higher, (c) a majority of the first few orders of the fundamental frequency must exist as peaks, typically 2 of the first 3 or 3 of the first 5 and so on.

[0018] In another aspect, the method may undertake cluster optimization by applying a method to identify order peaks occurring in more than one cluster, establishing by one or more rules or a probability-based model an assessment of the relationship of the peak to other peaks in the cluster, to which cluster it is most likely to belong, and removing it from the other clusters, such that, but not limited to, a) the specific fundamental frequency of the peak is closer to the mean of the fundamental frequency of one cluster than the other, b) the order of the peak in one cluster is lower than its order in the other.

[0019] In another aspect, the method may undertake sideband optimization by applying a method to establish whether each sideband peak is also a peak present in another group by one or more probability-based rules or model such that, but not limited to, (a) the peak is a sideband component in one group and an order component in the other (b) the peak is positioned at a lower sideband delta frequency multiple in one group than the other, (c) the base delta frequency of the sideband peak is closer to the average of one group than the other group, (d) the sideband peak is part of a larger cluster in one group than the other.

[0020] In another aspect, the method may also undertake sideband optimization by applying a method to establish that each sideband peak in a group is likely or not to be truly related to that group by one or more probability-based rules or model such as, but not limited to, (a) depending on whether its central peak is an order peak in the same group, (b) if delta frequencies of the sideband group are all multiples of the base delta frequency of 2, 3 or more, then the sidebands belong to another base delta frequency and if they are outside the defined sideband range or are not deemed likely, the sideband peaks are removed from the band.

[0021] According to another aspect, the method may undertake a selection of only the group having (a) the largest number of peaks and / or (b) the strictest cluster dimension (tolerance) and / or (c) the best pattern of harmonic peak presence, is selected as being representative of the most likely symptom related to that defined fault type and is stored as group component matrices.

[0022] According to another aspect, the method may undertake a selection of only the sideband group having (a) the largest number of peaks and / or (b) the strictest cluster dimension (tolerance) and / or (c) the best pattern of sideband peak presence, is selected as being representative of the most likely sidebands related to that specific defined defect type and is stored as matrices of sideband components associated with that specific group.

[0023] According to another aspect, the method may undertake an application of a noise sheet filter to filter noise from the frequency spectrum to eliminate unwanted spectral noise, which is achieved by keeping only the components located +10dB above the local spectral sheet level, and in which the remainder is reduced to zero.

[0024] According to another aspect, the method may undertake peak identification of the noise-filtered spectrum, then a quadratic interpolation process of the peaks is performed on the unfiltered spectrum, and it identifies the exact frequency as orders of presumed shaft speed and the amplitude of each possible peak and stores them in a matrix. The result creates normalized matrices of peak frequencies and peak amplitudes.

[0025] In another aspect, the method may undertake known mechanical frequency peak removal using a module leaving a set of frequencies including frequency components of bearing defects such as the inner and outer rings of the bearing. The mechanical peaks result from the test function such as shaft speed rather than the bearing, and wherein a "Funds in Band" sub-algorithm is called at least once in sequence to select matrix peaks that might have a sub-order and thus have a harmonic of a frequency within a specified frequency band.

[0026] According to another aspect, the method may undertake a determination with the "fundamentals in band" function whether each peak itself or an exact suborder thereof falls within the frequency range of fundamental frequencies of interest (FunL -> FunH) of a creation of a new matrix of fundamental frequencies, harmonic rank and original indices. Those whose fundamental frequencies are grouped within prespecified or learned tolerances are grouped as "orders" for that specific fundamental frequency. Then, a "Sides in Band" function is applied and for the selected group(s) for each order of that group, every possible peak in the original matrix is ​​checked to establish whether its delta frequency relative to that order is or has a suborder in the sideband range.

[0027] According to another aspect, the method may undertake a determination of the artificially learned tolerances.

[0028] In another aspect, the method may undertake an association with the clustering orders sub-algorithm of frequency elements that belong to a frequency group, wherein a dimensional clustering method performs group outlier removal, wherein the groupings by specific tolerances may be performed by (i) histograms of decreasing bin sizes for one or more groups, (ii) clustering of decreasing size by removing outliers for the same group, (iii) multiple clusters for multiple groups, and wherein the results of Clusterorders and / or Remove-Subs sub-algorithms are matrices of individual order components sorted into identifiable groups.

[0029] In another aspect, the method may cause the search range of the fundamental frequency to relate to a first iteration centered around the expected frequency of a known mechanical vibration component and using a low number of harmonics, then by clustering the results, the most plausible cluster is used to calculate a considerably narrower search range and the number of harmonics is increased for a second iteration, whereupon, with or without the option of further clustering, the identified peaks are removed from the peak matrices before the iteration of steps a), b), c), d), e), f) above for the fault frequencies.

[0030] According to another aspect, the method may undertake a side-ignoring test comprising a) the group has at least 2 "first" sideband components, and / or b) the group has at least 1 "first" sideband component AND 2 "second" sideband components. The results of Clus-terSides and / or IgnoreSides functions are matrices of individual sideband components sorted into the same group(s) of the order component that is their "center frequency.

[0031] According to another aspect, the method may undertake a weighting of each component in the model according to two sets of weighting coefficients; one for the harmonics (orders) and one for the statistically derived sidebands to determine the most likely fault type and these weightings being static algorithm parameters.

[0032] According to another aspect, the method may undertake the use of the "Most out of N" principle which is a probabilistic sub-algorithm to confirm with a high degree of confidence whether a bearing defect is present or not, in which the values / measurement parameters required for the N last measurements include 1) bearing defect frequency types, 2) a bearing defect frequency value from 1 to N, 3) a fundamental bearing defect frequency value from 1 to N, 4) a bearing defect sideband frequency value from 1 to N, and 5) revolutions / minute from 1 to N.

[0033] According to one or more non-limiting embodiments, additional features and advantages are achievable using the techniques of the present disclosure. Other embodiments and aspects of the disclosure are described in detail herein. For a better understanding of the disclosure and the advantages and features, reference is made to the description and drawings. Brief description of the figures

[0034] The above and other features and advantages of the embodiments of the present document will become apparent from the following detailed description, taken in conjunction with the accompanying drawings in which:

[0035] [Fig.l] represents a flowchart illustrating an exemplary basic algorithm in accordance with one or more embodiments of the present invention,

[0036] [Fig.2] represents a block diagram of a basic system for realizing the function of the proposed basic algorithm in accordance with one or more embodiments,

[0037] [Fig. 3] represents a flowchart of an application of the present invention, and

[0038] [Fig.4] represents a process flow of the process steps required for the rea lization of the function of the algorithm according to the present invention. Detailed description of the invention

[0039] Embodiments of the present document relate to a feature of a proposed method for automatically detecting bearing exfoliation defects on railway axlebox bearings without the need to know the exact bearing designation (or defect frequencies) or to have an accurate shaft speed. The method involves a "reverse scan" and the results of multiple identified patterns are disclosed. With the right parameters, the algorithm is applicable to railway monitoring systems and other industries without knowing the exact bearing designation (or defect frequencies) or without an accurate shaft speed. A sub-algorithm or module typically includes software that executes a specific instruction set.Those skilled in the art will recognize, however, that a specific instruction set may also be executed by hardware or a combination of software and hardware.

[0040] Throughout this document, the term "process" is used to describe the implementation implementation of various functions. A method may provide several different ways of performing the intended function and is not intended to be limiting.

[0041] [Fig.l] represents a flowchart illustrating an algorithm 100 that structures the functionality of a proposed method for the automatic detection of bearing exfoliation defects on railway axlebox bearings without the need to know the exact bearing designation or defect frequencies) or to have an accurate shaft speed. The automatic detection algorithm may be executed on a processor 102 (see [Fig.2]) or other computing structure comprising and / or using any number and combination of computing devices and networks using various communication technologies, as described herein. The computing structure may be readily scalable, extensible, and modular, with the ability to switch to different services or reconfigure certain functionalities independently of others.The algorithm may be a computer-executable program or a computer-executable product.

[0042] [Fig.2] shows a schematic diagram of a basic system 200 for performing the function of the proposed method which comprises a processor 102 and one or more sensors 104 connected to at least one bearing 105 mounted on a set of train wheels 108.

[0043] In a sub-algorithm module 110, condition monitoring data comprising vibration harmonics of the at least one bearing 105 coupled to a rotating shaft 109 is received by the processor 102. The vibration data, previously or sequentially transformed into the frequency domain, is processed into magnitude and frequency matrices. These comprise an approximate shaft speed in revolutions / minute or revolutions / second.

[0044] In a sub-algorithm module 120, a peak determination sub-algorithm module is applied. The peak determination sub-algorithm determines individual peaks from the background noise, as well as the actual amplitudes of the peaks and their exact peak center frequencies. The exact frequency and amplitude of each possible peak are stored in a matrix. Here, the frequency of each identified peak is divided by the frequencies of all other peaks in turn. The absolute value of the result is divided by the number 1 up to the number of sidebands to be searched for in order to identify all possible sideband peaks.

[0045] Regardless of the format in which the condition monitoring data is provided, it is often converted to an enveloped acceleration which is normally performed by an acquisition system and then to a spectrum preferably with peak-to-peak scaling and using the Hanning window before an FFT. The FFT results are converted to 'magnitude' and the spectrum is truncated into bins of numbers and with a spectral FMax of (waveform sampling rate sps) / 2.56. For example, a waveform with 4096 samples acquired at 2560 sps will produce a spectrum of 1601 bins (including the Zero bin) with an FMax of 1000 Hz (last bin).

[0046] To improve the identification of spectral peaks, it is advantageous to zero out all spectral components considered to be in or near the noise sheet and keep only those that are significantly above the noise sheet for the model scan correlation. A filter threshold of 10 dB and above gives good results in removing unwanted noise. Here, a module to calculate the local noise sheet level and remove the "noise" components from the noise-filtered spectrum is provided. This is achieved by scanning the spectrum bin by bin, calculating the median using a number of bins on each side of that specific bin corresponding to + and -1 / 32 of the number of spectral bins. If that specific bin has a value greater than or equal to the median plus 10 dB (or median*3.16), its value is kept otherwise its value is zero.

[0047] For each NON-ZERO bin value in the FILTERED noise spectrum, it is necessary to determine whether it lies within a bin spacing and is representative of a "Peak" and calculate its exact frequency and amplitude by the Peak Identification / Quadratic Interpolation method using the UNFILTERED spectrum to create matrices of (a) peak frequencies (PkHz) and (b) peak amplitudes (PkMags).

[0048] In a sub-algorithm module 130, all possible integer sub-orders of each peak up to a predefined number of orders including the peak itself that fall within a specific range of target fundamental frequency ranges are identified. Here, each peak frequency is divided by a number from 1 up to the number of harmonics to be included in turn to determine all possible theoretical sub-fundamental frequencies. One for each defined fault type is identified. The peaks are stored with their harmonic rank and theoretical fundamental frequency as order matrices.

[0049] The parameters regarding the fault characteristics 135 are stored in a separate database and can be accessed by the sub-algorithm module 130. The fault information is related to each defined fault type, such as BPFO Bail (roller) Pass Frequency Outer-race (roller) and BPFI Bail (roller) Pass Frequency Inner-race (roller)

[0050] The fault fundamental frequency as an order of operating speed is specific to the fault. Examples may be 6.23 and 11.34, etc. A typical fault fundamental frequency range / characteristic in + / -% is <1% at 33 %. A default / characteristic fault sideband base delta frequency as an operating speed order is 037, 1.0, etc. Further, a range of default / characteristic fault sidebands in + / - % is <1 to 33. Further, a range of default / characteristic fundamental frequency (harmonic) orders to be included is 3 to 100+. Finally, a number of default / characteristic sidebands (on each side) to be included is 0 to 10. Thus, all fault information disclosed in this paragraph is accessible by the sub-algorithm module 130 to identify faults.

[0051] In a sub-algorithm module 140, all peaks in the original list that may be possible sideband components of the peaks identified in the sub-algorithm module 130 are identified. The peaks that could be possible sideband components are identified by those having a base delta frequency within a specific range of target sideband frequency ranges, one for each defined sideband type, and up to + / - delta orders of a predefined sideband to determine all possible theoretical base delta sideband frequencies and store these possible sideband peaks with the sideband number. The theoretical delta sideband frequency and the center peak ID index, which are within each of the target sideband ranges, are stored as sideband matrices.All the characteristics relating to defects 135 described in this paragraph can be consulted by the sub-algorithm module 140 in order to identify possible peaks.

[0052] In a sub-algorithm module 145, a mechanical suppression function is performed to eliminate mechanical noise. This option inserted between the Peak Matrices module 120 and the fundamental frequency peak processing modules 130 and 140 is used to suppress wheel noise components. Only the group having (a) the largest number of peaks and / or (b) a strictest cluster dimension (tolerance) and / or (c) the best pattern of harmonic peak presence, is selected as being representative of the most likely symptom related to that defined fault type and is stored as group component matrices. The mechanical suppression function can be manually "on" or "off" or even configured to depend on the amount of wheel noise activity present.

[0053] The mechanical suppression function option 145 obtains its inputs from databases that contain parameters and characteristics common to all mechanical noises. A first input is received from a mechanical clustering parameter module 147. In the mechanical clustering parameter module 147, a default / characteristic range for the smallest acceptable cluster size in number of peaks is in a range between 2 and 5. The limit for clustering Mechanical order clustering as execution speed orders for a first iteration is approximately 0.01. Note: This default / characteristic range may vary in value and definition depending on the clustering method used since many types of clustering methods can be used such as DBSCAN, OPTICS, Gaussian Mixture, etc. The limits for clustering mechanical orders as operation speed orders for a second iteration are for a "distance between nearest points" approximately 0.006. The default / characteristic range for the second iteration may also vary in value and definition depending on the clustering method used. The limit for clustering mechanical sidebands as operation speed orders is approximately 0.01. The default / characteristic range may vary in value and definition depending on the clustering method used.

[0054] A second input to the mechanical suppression function 145 is received from a second mechanical suppression module 149. Here, a characteristic default mechanical fundamental frequency as an operating speed or shaft speed order is 1. A default / characteristic fault fundamental frequency range in + / - % is <1% to 33%. A default / characteristic fault sideband base delta frequency as an operating speed order is 037, 1.0, etc. Further, a default / characteristic fault sideband range in + / - % is <1 to 33. A default / characteristic fundamental frequency harmonic range to be included in a first iteration is between 3 and 11. Further, a default / characteristic fundamental frequency harmonic range to be included in a second iteration is 0 or up to 100+.Finally, a default / characteristic number of sidebands (on each side) to be included is 0 to 10. Thus, all of the defect information disclosed in this paragraph is accessible by the mechanical suppression function 145 in order to identify mechanical defects.

[0055] In a sub-algorithm module 150, peaks having theoretical fundamental frequencies within each specific target range are clustered by one or more of numerous methods. Theoretical fundamental frequencies having a predefined cluster size limit and clusters having more than a predefined number of peaks are organized into groups and all others as outliers, and the groups are stored in order matrices. A "group" in a measurement consists of spectral components (identified peaks) related either because they are almost real orders of a fundamental frequency within the fundamental frequency range of interest being that of one of the peaks, or theoretical and other spectral components having all delta frequency up to an "order" being almost exactly the same and within the sideband range of interest.

[0056] In a sub-algorithm module 155, fault clustering parameters common to all defined faults are stored in a separate database and can be consulted by the sub-algorithm module 150. A smallest acceptable cluster size in number of peaks is in a range between 2 and 5. Limits of fault order clustering as orders of operating speed are approximately 0.005. Limits of fault sideband clustering as orders of operating speed are approximately 0.01. All characteristics regarding fault clustering parameters 155 disclosed in this paragraph can be consulted by the sub-algorithm module 150 to identify possible peaks.

[0057] In a sub-algorithm module 157, a decision between transferring the group orders may or may not be optimized. If the group orders are to be optimized in the sub-algorithm module 159, only those groups having (a) the largest number of peaks and / or (b) the strictest cluster dimension (tolerance) and / or (c) the best pattern of harmonic peak presence, are selected as being representative of the most likely symptom related to that defined fault type and are stored as group component matrices.

[0058] During optimization in the sub-algorithm module 159, unwanted groups are removed from a list of potential groups stored in the matrices. The groups to be removed are typically a subgroup of another group having a lower or higher fundamental frequency, such as, but not limited to, (a) the majority of the order peaks in the group are present as orders in another group of lower fundamental frequency, (b) the orders of the peaks are all multiples of 2, 3 or higher, (c) a majority of the first few orders of the fundamental frequency must exist as peaks, typically 2 of the first 3 or 3 of the first 5, etc.

[0059] Additionally, during the optimization of groups 159, the order peaks that are present in more than one group are identified. This is determined by one or more rules or a probability-based model that evaluates the relationship of the peak with the other peaks in the group, the group to which it is most likely to belong, and removes it from the other groups. This includes, but is not limited to, if (a) the specific fundamental frequency of the peak is closer to the average of a fundamental frequency of a group than the other, if (b) the order of the peak in one group is lower than its order in the other.

[0060] The sub-algorithm module 160 determines those of the sidebands of the sideband matrices that are related to the central frequency components present in a group component matrix and have sideband delta frequencies within the target sideband frequency range for this definition of group fault. A clustering sub-algorithm is applied to their basic delta frequencies with a predefined cluster size limit. Clusters having more than a predefined number of sideband peaks are organized into sideband groups and all others as outliers. The sideband groups are stored in matrices with references to their respective center frequency peak from the order matrices. All features regarding the fault clustering parameters 155 described above, which are stored in a separate database, are accessible by the sub-algorithm module 160 to identify possible peaks.

[0061] In a sub-algorithm module 167, a decision between transferring the sidebands of the groups may or may not be optimized. If the sidebands of the groups are to be optimized in the sub-algorithm module 169, only the sideband group having (a) the largest number of peaks and / or (b) the tightest cluster size (tolerance) and / or (c) the best pattern of sideband peak presence, is selected as being representative of the most likely sidebands related to that specific defined defect type and is stored as matrices of sideband components associated with that specific group.

[0062] During optimization 169, it is determined whether each sideband peak is also a peak present in another group by one or more rules or a probability-based model. The rules include, but are not limited to, (a) the peak is a sideband component in one group and an order component in the other (b) the peak is positioned at a multiple of a lower sideband delta frequency in one group and not in the other, (c) the base sideband delta frequency of the peak is closer to the mean of one group than the other group, d) the sideband peak is part of a larger cluster in one group rather than the other.

[0063] Further, during sideband optimization, each sideband peak in a group is determined to be likely or unlikely to be truly related to that group by one or more probability-based rules or model. The rules include, but are not limited to, (a) whether its center peak is an order peak in the same group, (b) if delta frequencies of the sideband group are all multiples of the base delta frequency by 2, 3 or more, then the sidebands belong to another base delta frequency and if they are not within the defined sideband range or deemed not likely, the sideband peaks are removed from the group.

[0064] After both the group orders are optimized 159 and the group sidebands are optimized 169 (or not) in their respective sub-algorithm modules, the two outputs are combined into a single output containing the groups combined represented in a sub-algorithm module 175. The output of the combined optimized groups provides a person skilled in the art with the information necessary to automatically detect bearing defects. In other words, the basic algorithm allows, so to speak, to separate the wheat from the chaff by providing models of related spectral components from each defined range of interest. At this stage of the disclosure, the essential function of the present invention has been described.

[0065] Certain applications of the basic algorithm 100 will now be disclosed. [Fig.3] represents an application 300 of the basic algorithm 100. [Fig.3] illustrates the introduction of the result of the basic algorithm into the sub - algorithm module 310. In the sub - algorithm module 310, the model correlation values are grouped.

[0066] In the module, the model weightings are introduced into the sub - algorithm module 310. Here, for each rolling defect and for each mechanical symptom, a model is created with a fundamental frequency component, no or some harmonic components (orders 2, 3, etc. of the fundamental frequency) and no or some sideband components having a delta frequency other than its fundamental frequency (generally another fundamental frequency). For the rolling models, three factors must be defined;

[0067] harmonics as the number of orders (including the fundamental frequency);

[0068] The number of sidebands; and

[0069] The defect frequency or the mechanical frequency identifier for the sidebands.

[0070] The individual weightings of the models are derived by: component weighting = order # weighting * sideband # weighting. Thus, the 3rd order of the 2nd sideband of the fundamental frequency would be = 0.8 * 0.9 = 0.72 and so on. Then, depending on the number of harmonics and / or sidebands, the model weightings are normalized so as not to favor models with more components by;

[0071] Normalized Weightings = (Weightings in the model) / SQRT(RSS(Weightings in the model)). Here, RSS is the value of the sum of the square roots.

[0072] Each component in the model is weighted according to two sets of weighting coefficients, one for harmonics (orders) and the other for sidebands. These coefficients remain constant regardless of the application.

[0073] In the sub-algorithm module 320, a system based on the Moon principle (making the most of N) is used to determine whether a bearing defect is present or not. This requires that at least the N most recent measurements be kept in a FIFO buffer OR that all are kept in a database that the module's logic function can access.

[0074] In sub-algorithm module 330, the Most-out-of-N (MooN) logic concept considers at least 3 levels of conditions that must be satisfied to prove fault symptom consistency in most of the last N measurements before declaring a detection. For many applications with random noise presence, this MooN method, whether applied or not for an alarm hysteresis rules-logic approach, has proven invaluable in achieving decent detection reliability.

[0075] One approach is hard coding of MooN logic or thresholds such as;

[0076] Presence of orders in the first three orders,

[0077] Presence of lateral bands;

[0078] rss of group components or pob (band percentage = rss(group) / rss(band)*100);

[0079] Coherence of the fundamental frequency between sequential measurements

[0080] Consistency of sidebands present or not between sequential measurements

[0081] Consistency of sideband delta frequency between sequential measurements.

[0082] In the last sub-algorithm module 340, a RatioStats function is considered to determine how many of the frequency ratios provided for both the fundamental frequencies and the sidebands (if any) have a statistical distribution within the RFthresh limits. First, the fundamental ratios are evaluated, then of those falling below the threshold limits, their sideband frequency ratios if present are evaluated. Various evaluation options have been tested and all are valid, but they have not been tested on enough data to establish which is the most reliable.

[0083] Option A; Range + / - from mean as a percentage of mean

[0084] Option B; Range + / - from median as a percentage of median

[0085] Option C; Clustering methods such as DBSCAN, OPTICS, mixing Gaussian, etc.

[0086] Option D; A statistical value (i.e., standard deviation), etc. For the purposes of this document and as it was the initial starting point, Option A is described below. The methodology is to first evaluate the fundamental frequency ratio ranges of all samples provided, if the result is outside the provided limit, then remove the "outlier" and re-evaluate and repeat until the result is within the limit, if this number of measurements is greater than N / 2, then for these measurements only (and if sidebands are present) repeat for the sideband frequency ratios. The number of measurements (whether the original number, a reduced number, or even 1) that meet to the limits is returned

[0087] Based on the RatioStats, model probability calculation (SPPC) results are determined based on a binary value of each of the defined bearing defects to consider them as present (1 = detected) or not (0) with a significant degree of confidence compared to the standard condition monitoring alarms. It will be noted that only the most likely defects present will have a detection logic of 1.

[0088] A method of performing automatic bearing fault detection 400 will now be disclosed with reference to FIGS. 1 to 4.

[0089] In a first step 410, the method undertakes a reception, by a processor from one or more sensors, of condition monitoring data. The condition monitoring data comprises harmonics of vibrations of at least one bearing coupled to a rotating shaft. The bearing having an inner ring and an outer ring.

[0090] In a step 420, the method undertakes a reception of vibration data transformed previously or sequentially into the frequency domain, which are finally provided in the form of matrices of quantities and frequencies, with an approximate shaft speed in revolutions / minute or revolutions / second.

[0091] In a step 430, the method undertakes an application of a peak determination method that determines individual peaks from the background noise, as well as the actual amplitudes of the peaks and their exact peak center frequencies. The exact frequency and amplitude of each possible peak are stored in a matrix.

[0092] In a step 440, the method undertakes an identification of all possible integer sub-orders of each peak up to a predefined order number, including the peak itself, which fall within a specific range of target fundamental frequency ranges, one for each defined defect type and the storage of these peaks with their harmonic rank and theoretical fundamental frequency as order matrices.

[0093] In a step 450, the method undertakes an identification from the original list of all peaks that could be possible sideband components of the peaks identified in step 440, having a base delta frequency in a specific range of target sideband frequency ranges, one for each defined defect type, and up to + / - delta orders of a predefined sideband to determine all possible theoretical base delta sideband frequencies and a storage of these possible sideband peaks with the sideband number, the theoretical delta sideband frequency and the center peak ID index, which fall in each of the target sideband ranges as sideband matrices.

[0094] In a step 460, the method undertakes an identification of peaks having theoretical fundamental frequencies included in each specific target range and which are clustered by one or more of numerous methods. Theoretical fundamental frequencies with a predefined clustering size limit and clusters containing more than a predefined number of peaks are organized into groups and all others as outliers. The organized groups are stored as order matrices.

[0095] In a step 460, the method undertakes a determination, from the sideband matrices, of the sidebands which are related to center frequency components present in a group component matrix and having delta sideband frequencies in the target sideband frequency range for this group defect definition. A clustering process is applied to their basic delta frequencies with a predefined cluster size limit and clusters containing more than a predefined number of sideband peaks are organized into sideband groups and all others as outliers. The sideband groups are stored in matrices with references to their respective center frequency peak from order matrices.

[0096] In a step 470, the method undertakes model clustering for each bearing inner or outer ring defect (or other defect types). Three model building factors are defined by 1) harmonic identification as number of orders, 2) number of sidebands and 3) fundamental frequency and mechanical frequency identifier for the sidebands, wherein one fundamental frequency component, none or some harmonic components, none or some sideband components having a delta frequency other than its fundamental frequency compose the possible clustered models.

[0097] In a step 480, the method undertakes a determination of state indicators. Here, the most plausible group can then be selected by (a) correlation with weighted model components, (b) presence or absence of components (c) RSS of group components (d) pattern recognition method on each group.

[0098] In a step 490, the method undertakes a fault identification with a high level of confidence. Faults can thus be predicted with a high degree of confidence by a) the Moon principle; that is, for each logic test, most of the results of a test set of N measurements must be "positive", b) by hardcoding, in which the algorithm is coded using a higher-level compiler or computer resource La, c) by machine learning in conjunction with the use of an artificial intelligence algorithm to "learn" the faults that exist.

[0099] The method further undertakes a conversion of the magnitude and frequency matrices produced in step 410 into an acceleration wrapped by a data acquisition system, and / or a Hanning window is performed on the data before the TFR, and / or a TFR or a TFRep is performed on data to generate spectral data, and / or the magnitude of the TFR data is calculated from the real and imaginary spectral components, "Magnitude" = SQRT(RealA2 + hnagA2), and / or the spectrum is truncated.

[0100] The method undertakes a division of the peak frequency produced in step 430 by a number from 1 to the number of harmonics to be included in turn to determine all possible theoretical sub-fundamental frequencies.

[0101] The method further undertakes a subtraction of the frequency of each peak identified in step 420 by the frequency of all other peaks and a division of the absolute value of the result by the number 1 up to the number of sidebands to be searched in order to identify all possible sideband peaks.

[0102] The method further undertakes optimization of the groups determined in step 460 by removing from a list of groups those groups that are a subgroup of another group having a lower or higher fundamental frequency, such that, but not limited to, (a) the majority of the order peaks in the group are present as orders in another group of lower fundamental frequency, (b) the orders of the peaks are all multiples of 2, 3 or more, (c) a majority of the first few orders of the fundamental frequency must exist as peaks, typically 2 of the first 3 or 3 of the first 5 and so on.

[0103] The method further undertakes cluster optimization by identifying order-organized peaks from step 450 that are in more than one cluster. This is established by one or more probability-based rules or models that evaluate the relationship of the peak to other peaks in the cluster, the cluster to which it is most likely to belong, and remove it from the other clusters. The probability-based rules or models include: (a) the specified fundamental frequency of the peak is closer to the average of a fundamental frequency of one cluster than the other, (b) the order of the peak in one cluster is lower than its order in the other.

[0104] The method further undertakes a sides optimization by establishing whether each sideband peak is also a peak present in another group by one or more probability-based rules or model such that, but not limited to, (a) the peak is a sideband component in one group and an order component in the other (b) the peak is positioned at a lower sideband delta frequency multiple in one group than the other, (c) the base delta frequency of the sideband peak is closer to the average of one group than the other group, (d) the sideband peak is part of a larger cluster in a group than in the other.

[0105] The method undertakes sideband optimization by establishing whether each sideband peak in a group is likely to be truly related to that group by one or more probability-based rules or patterns such as, but not limited to, (a) whether its central peak is an order peak in the same group, b) if delta frequencies of the sideband group are all multiples of the base delta frequency of 2, 3 or more, then the sidebands belong to another base delta frequency and, if they are outside the defined sideband range or are not deemed likely, the sideband peaks are removed from the group.

[0106] The method further undertakes a selection of only the group from step 450 having (a) the largest number of peaks and / or (b) the strictest cluster dimension (tolerance) and / or (c) the best pattern of harmonic peak presence, as the most likely symptom related to that defined fault type and is stored as group component matrices.

[0107] The method undertakes a selection of only the sideband group from step 460 having (a) the largest number of peaks and / or (b) the strictest cluster dimension (tolerance) and / or (c) the best pattern of sideband peak presence as being representative of the most likely sidebands related to that specific defined defect type and is stored as matrices of sideband components associated with that specific group.

[0108] The method comprises applying a noise sheet filter to filter noise from the frequency spectrum to remove unwanted spectral noise by keeping only components located +10dB above the local spectral level of the sheet. All others are reduced to zero.

[0109] The method further undertakes peak identification of the noise-filtered spectrum and then quadratic interpolation of the peaks onto the unfiltered spectrum. This identifies the exact frequency as orders of assumed shaft speed and amplitude of each possible peak and then stores these in a matrix. The result creates normalized matrices of peak frequencies and peak amplitudes.

[0110] The method includes removing known mechanical frequency peaks using a module leaving a set of frequencies including the frequency components of bearing defects such as the inner and outer rings of the bearing. The mechanical peaks result from a test function such as shaft speed rather than the bearing. A "fundamentals in band" sub-algorithm is called twice to select matrix peaks that might have a sub-order and thus a theoretical fundamental frequency within a specified frequency band.

[0111] The method undertakes a determination with an algorithm of the fundamentals in the band if each peak itself or an exact suborder thereof lies within the frequency range of fundamental frequencies of interest (FunL -> FunH) of a creation of a new matrix of fundamental frequencies, harmonic rank and original indices. Those whose fundamental frequencies are clustered within a predefined or learned tolerance are grouped as "orders" for that specific fundamental frequency. Then, a "Sides in Band" function is applied and for the selected group(s) for each order in that group, each possible peak in the original matrix is ​​checked to establish that its delta frequency relative to that order is or has a suborder in the sideband range.

[0112] The method undertakes an association of the frequency elements that belong to a frequency group with a sub-algorithm of clustering orders.Here, a dimensional clustering method undertakes outlier removal from the group. Groupings by specific tolerances can be achieved by (i) histograms with decreasing bin sizes for one or more groups, (ii) clustering of decreasing size by removing outliers from a single group, (iii) multiple clusters for multiple groups. The outputs of Clusterorders and / or Remove-Subs sub-algorithms are matrices of individual order components sorted into identifiable groups.

[0113] The method provides that a search range for the fundamental frequency relates to a first iteration centered around the expected frequency of a known mechanical vibration component and using a low number of harmonics. Then, by clustering the results, the most plausible cluster is used to calculate a significantly narrower search range and the number of harmonics is increased for a second iteration, from which, with or without the possibility of further clustering, the identified peaks are removed from the peak matrices before iterating claim 1 for the fault frequencies.

[0114] The method further undertakes a test for a side-ignoring function which comprises:

[0115] a.) the group comprises at least 2 "first" sideband components, and / Or

[0116] b.) the group comprises at least 1 "first" sideband component AND 2 "second" sideband components. Here, the results of Clus-terSides and / or IgnoreSides functions are matrices of individual sideband components sorted into the same group(s) of the order component that is their "center frequency.

[0117] The method also includes weighting each component produced in step 470 in the model according to two sets of weighting coefficients; a for harmonics (orders) and one for statistically derived sidebands to determine the most likely fault type. These weights are static algorithm parameters.

[0118] The method provides confirmation with a high degree of confidence that a bearing defect is present or not according to the "Most-out-of-N" principle which is a probabilistic sub-algorithm that a bearing defect is present or not. The measurement values / parameters required for the last N measurements include:

[0119] bearing fault frequency types,

[0120] a bearing fault frequency value of 1 to N,

[0121] a fundamental bearing fault frequency value of 1 to N,

[0122] a rolling fault sideband frequency value of 1 to N, and

[0123] revolutions / minute from 1 to N.

[0124] The present invention may be a system, a method, and / or a computer program product at any possible level of technical integration. The computer program product may include a computer-readable storage medium (or media) having computer-readable program instructions enabling a processor to execute aspects of the present invention.

[0125] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, may be implemented by computer-readable program instructions.

[0126] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine, such that the instructions, executed through the processor of the computer or other programmable data processing device, create means for implementing the functions / actions specified in the flowchart block(s) and / or schematic diagram(s).These computer-readable program instructions may also be stored in a computer-readable storage medium and cause a computer, a programmable information processing apparatus and / or other devices to operate in a particular manner, such that the computer-readable storage medium on which instructions are stored includes an article of manufacture having instructions that implement aspects of the function / action specified in the flowchart block(s) and / or schematic diagram(s).

[0127] The computer-readable program instructions may also be loaded onto a computer, other programmable information processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus, or other device to produce a computer-implemented process, such that the instructions executed on the computer, other programmable apparatus, or other device implement the functions / actions specified in the flowchart and / or block diagram block(s).

[0128] The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which includes one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the blocks may proceed other than in the order shown in the figures. For example, two successive blocks may, in fact, be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order, depending on the function involved.It will also be noted that each block of the block diagrams and / or flowchart illustration, as well as combinations of blocks in the block diagrams and / or flowchart illustration, may be implemented by special-purpose hardware systems that perform the specified functions or actions or undertake combinations of special-purpose hardware and computer instructions.

[0129] The terminology used herein is intended to describe particular embodiments only and is not intended to be limiting. As used herein, the singular forms "a," "an," "the," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It is further understood that the terms "includes" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of any other entities, integers, steps, operations, components of elements, and / or groups thereof.

[0130] The descriptions of the various embodiments herein have been presented for purposes of illustration but are not intended to be exhaustive or limited to the disclosed embodiments. Numerous modifications and variations will occur to those of ordinary skill in the art without departing from the scope and spirit of the disclosed embodiments. The terminology used herein has been chosen to best explain the principles of the embodiments, the practical application or technical improvement over technologies found on the market, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

Claims

1. A method for automatically detecting bearing defects (105) comprising: receiving, by a processor (102) from one or more sensors (104), condition monitoring data, the condition monitoring data comprising harmonics of vibrations of at least one bearing (105) coupled to a rotating shaft (109), the bearing (105) having an inner ring and an outer ring, the method further comprising; a) receiving vibration data previously or sequentially transformed into the frequency domain and finally provided as magnitude and frequency matrices, with an approximate shaft speed in revolutions / minute or revolutions / second, b) applying a peak determination method which determines from the background noise of the individual peaks, the actual peak amplitudes and their exact peak center frequencies, the exact frequency and amplitude of each possible peak being stored in a matrix, (c) identifying all possible integer suborders of each peak up to a predefined order number, including the peak itself, which fall within a specific range of target fundamental frequency ranges, one for each defined defect type and storing these peaks with their harmonic rank and theoretical fundamental frequency as order matrices, (d) clustering of theoretical fundamental frequencies that fall within each specific target range, where theoretical fundamental frequencies have a predefined cluster size limit and clusters with acceptable clustering characteristics by numbers of peaks they contain and / or the strictest cluster sizes are organized into groups and all others as outliers, and the groups are stored in order matrices, and e) identifying all peaks that could be possible sideband components of the peaks identified in step (c), having a base delta frequency within a specific range of target sideband frequency ranges, one for each defined defect type, and up to + / - delta orders of a predefined sideband to determine all possible theoretical basic delta sideband frequencies and storing these possible sideband peaks with the sideband number, theoretical delta sideband frequency and center peak ID index, which fall into each of the target sideband ranges as sideband matrices, each basic delta frequency being a predetermined frequency associated with a defect according to the order of the bearing rolling speed, and f) applying a clustering method to their basic delta frequencies with a predefined cluster size limit and the clusters having acceptable clustering characteristics by the number of peaks they contain and / or the most stringent cluster sizes are organized into sideband groups and all others as outliers,and the sideband groups are stored in matrices with references to their respective center frequency peaks from the order matrices.,

2. A method for automatic bearing defect detection according to claim 1, further comprising grouping models for each inner or outer ring bearing defect (or other types of defects), wherein three factors for model construction are defined by 1) identification of harmonics as number of orders, 2) number of sidebands and 3) fundamental frequency range and frequency range for sidebands, wherein one fundamental component, none or few harmonic components, none or few sideband components having a delta frequency other than its fundamental frequency compose the possible grouped models.

3. A method for automatically detecting bearing defects according to claim 1, further comprising determining status indicators, wherein the selection of a most plausible group can then be performed by (a) correlation with weighted model components, (b) presence or absence of components (c) RSS of group components (d) pattern recognition method on each group.

4. A method for automatically detecting bearing defects according to claim 1, further comprising identifying defects with a high confidence level, wherein the defects can be predicted with a high degree of confidence according to the "Most-out-of-N" (MooN) principle; i.e., for each test logically, most results of a test set of N measurements should be "positive", with the MooN logical parameters being defined by a) experienced analysts, b) design of experiments (DOE) on known data, c) machine learning, in which an artificial intelligence algorithm to "learn" the parameters is used to identify defects that exist.

5. The method of claim 1, wherein the magnitude and frequency matrices are converted into a wrapped acceleration by a data acquisition system, and / or a Hanning window is performed on the data before a TFR, and / or a TFR or TFRep is performed on data to generate spectral data, and / or the magnitude of the TFR data is calculated from the real and imaginary spectral components, "Magnitude" = SQRT(ReelA2 + hnagA2), and / or the spectrum is truncated, Real and Imag being respectively real and imaginary spectral components obtained by the TFR.

6. The method of claim 1, further comprising dividing each peak frequency in step c) by a number from 1 up to the number of harmonics to be included in turn to determine all possible theoretical sub-fundamental frequencies.

7. The method of claim 1, further comprising subtracting each peak frequency identified in step b) by all other peak frequencies in turn and dividing the absolute value of the result by the number 1 up to the number of sidebands to be searched for to identify all possible sideband peaks.

8. The method of claim 1, further comprising optimizing groups by applying a method to remove from the list of groups groups that are a subgroup of another group having a lower or higher fundamental frequency, such that, but not limited to, (a) the majority of the order peaks in the group are present as orders in another group of lower fundamental frequency, (b) the orders of the peaks are all multiples of 2, 3 or more, (c) a majority of the first few orders of the fundamental frequency must exist as peaks, typically 2 of the first 3 or 3 of the first 5 and so on.

9. The method of claim 1, further comprising optimizing groups by applying a method of identifying order peaks lying in more than one group, established by one or more rules or a model based on the probabilities of a peak belonging to a peak of group that evaluate the peak's relationship to other peaks in the group, the group to which it is most likely to belong and remove it from other groups, such that, but not limited to, a) the specific fundamental frequency of the peak is closer to the mean of the fundamental frequency of one group than the other, b) the order of the peak in one group is lower than its order in the other.

10. The method of claim 1, further comprising optimizing sides by applying a method to establish whether each sideband peak is also a peak present in another group by one or more rules or a model based on the probabilities of a peak belonging to a group peaks such that, but not limited to, (a) the peak is a sideband component in one group and an order component in the other (b) the peak is positioned at a lower sideband delta frequency multiple in one group than the other, (c) the base delta frequency of the sideband peak is closer to the average of one group than the other group, (d) the sideband peak is part of a larger cluster in one group than the other.

11. The method of claim 1, further comprising optimizing sides by applying a method to establish whether each sideband peak in a group is likely to be truly related to that group by one or more rules or a model based on the probabilities of a peak belonging to a group peaks such as, but not limited to, (a) whether its central peak is an order peak in the same group, (b) if delta frequencies of the sideband group are all multiples of the base delta frequency of 2, 3 or more, then the sidebands belong to another base delta frequency and if they are outside the defined sideband range or are not deemed likely, the sideband peaks are removed from the group.

12. The method of claim 1, further comprising selecting only the group having (a) the largest number of peaks and / or (b) the strictest cluster size (tolerance) and / or (c) the best harmonic peak presence pattern, is selected as being representative of the most likely symptom related to that defined fault type and is stored as group component matrices

13. The method of claim 1, further comprising selecting only the group of sidebands having (a) the largest number of peaks and / or (b) the smallest cluster size (tolerance). strict and / or (c) the best sideband peak presence model, is selected as representative of the most likely sidebands related to that specific defined defect type and is stored as matrices of sideband components associated with that specific group.

14. The method of claim 1, further comprising applying a noise sheet filter to filter noise from the frequency spectrum to remove unwanted spectral noise is performed by keeping only components located +10dB above a local spectral sheet level, and wherein the remainder is reduced to zero.

15. The method of claim 14, further comprising identifying peaks in the noise-filtered spectrum and then performing a quadratic peak interpolation process on the unfiltered spectrum, which identifies the exact frequency as orders of assumed shaft speed and the amplitude of each possible peak and stores them in a matrix. The result creates normalized matrices of peak frequencies and peak amplitudes.

16. The method of claim 1, further comprising removing known mechanical frequency peaks using a module leaving a set of frequencies including frequency components of bearing defects such as the inner and outer rings of the bearing, wherein the mechanical peaks result from the test function such as shaft speed rather than the bearing, and wherein a "Funds in Band" sub-algorithm is called at least once in sequence to select matrix peaks that might have a sub-order and thus have a harmonic of a frequency within a specified frequency band.

17. A method according to claim 16, further comprising with the "fundamentals in band" function if each peak itself or an exact suborder thereof falls within the frequency range of fundamental frequencies of interest (FunL -> FunH) creating a new matrix of fundamental frequencies, harmonic rank and original indices, in which those whose fundamental frequencies are grouped within a predefined or learned tolerance are grouped as "orders" for that specific fundamental frequency, and then applying an "sides in band" function to the selected group(s) for each order in that group to check each possible peak in the original matrix to establish whether its delta frequency re- relative to this order is or has a suborder in the sideband range.

18. The method of claim 1, further comprising determining the tolerances learned by artificial intelligence.

19. A method according to claim 16, wherein a search range of the fundamental frequency relates to a first iteration centered around the expected frequency of a known mechanical vibration component and using a low number of harmonics, then by clustering the results the most plausible cluster is used to calculate a considerably narrower search range and the number of harmonics is increased for a second iteration, whereupon, with or without the option of further clustering, the identified peaks are removed from the peak matrices before the iteration of claim 1 for the fault frequencies.

20. A method according to claim 3, wherein each component in the model is weighted according to two sets of weighting coefficients; one for harmonics (orders) and one for statistically derived sidebands to determine the most likely fault type and wherein these weights are static algorithm parameters.

21. The method of claim 4, wherein the "Most out of N" principle is a probabilistic sub-algorithm for confirming with a high degree of confidence whether a bearing defect is present or not, wherein the measurement values / parameters required for the last N measurements include 1) bearing defect frequency types, 2) a bearing defect frequency value 1 to N, 3) a bearing defect fundamental frequency value 1 to N, 4) a bearing defect sideband frequency value 1 to N, and 5) RPM 1 to N.

Images