Bearing fault diagnosis method and device, storage medium and computer equipment

Abstract

The present disclosure provides a bearing fault diagnosis method, device, storage medium and computer equipment. The bearing fault diagnosis method comprises: obtaining a set of original vibration data of a target bearing; processing the set of original vibration data by each of a plurality of preset diagnosis models to determine whether each original vibration data in the set of original vibration data is fault data, and to calculate a fault data proportion of the set of original vibration data, wherein each preset diagnosis model corresponds to a fault data proportion; obtaining a weight of each of the plurality of preset diagnosis models; and determining a fault diagnosis result of the target bearing according to the weight of each of the plurality of preset diagnosis models and the corresponding fault data proportion. The present disclosure can obtain more accurate diagnosis results, thereby improving the precision and recall rate, helping to improve the generalization of the method, reducing the false positive rate and the false negative rate, and ensuring the reliable diagnosis of bearing faults.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of wind power generation, and more particularly, to a bearing fault diagnosis method and device, a storage medium and a computer equipment. BACKGROUND

[0002] In order to realize the fault diagnosis of rolling bearing vibration signals, there are various diagnostic models under different technical routes in the related art. However, no matter which model is used, due to the different characteristics of different bearings and the differences in the application effect of the model under various working conditions, there is a problem that a single method has poor generalization in actual industrial environment, and has high false positive rate and false negative rate. SUMMARY

[0003] Therefore, how to improve the generalization of the fault diagnosis method and reduce the false positive rate and the false negative rate is crucial for reliable diagnosis of bearing faults.

[0004] In one general aspect, there is provided a bearing fault diagnosis method, including: obtaining a set of original vibration data of a target bearing; processing the set of original vibration data by each of a plurality of preset diagnosis models to determine whether each original vibration data in the set of original vibration data is fault data, and to calculate a fault data proportion of the set of original vibration data, wherein each preset diagnosis model corresponds to a fault data proportion; obtaining a weight of each of the plurality of preset diagnosis models; and determining a fault diagnosis result of the target bearing according to the weight and the corresponding fault data proportion of each of the plurality of preset diagnosis models.

[0005] Optionally, the number of the target bearings is at least one, and the processing of the set of original vibration data by each of the plurality of preset diagnosis models to determine whether each original vibration data in the set of original vibration data is fault data and to calculate a fault data proportion of the set of original vibration data includes: processing the set of original vibration data of each of the target bearings by each of the plurality of preset diagnosis models respectively to determine whether each original vibration data in the set of original vibration data of each of the target bearings is fault data and to calculate a fault data proportion of the set of original vibration data of each of the target bearings.

[0006] Optionally, the bearing fault diagnosis method further comprises: determining, for each preset diagnosis model, a statistical value of a fault data proportion of the original vibration data set of each target bearing; determining, according to the statistical value of the fault data proportion corresponding to each preset diagnosis model, a fault data proportion threshold of each preset diagnosis model; and wherein the determining of the fault diagnosis result of the target bearing according to the weight of each preset diagnosis model in the plurality of preset diagnosis models and the corresponding fault data proportion comprises: determining the fault diagnosis result of the target bearing according to the weight, the corresponding fault data proportion and the fault data proportion threshold of each preset diagnosis model in the plurality of preset diagnosis models.

[0007] Optionally, the determining of the fault diagnosis result of the target bearing according to the weight, the corresponding fault data proportion and the fault data proportion threshold of each preset diagnosis model in the plurality of preset diagnosis models comprises: determining, for each preset diagnosis model, a ratio of the fault data proportion of the target bearing to the fault data proportion threshold, to obtain a preliminary diagnosis result of the target bearing under each preset diagnosis model; and determining a weighted average value of the preliminary diagnosis results of the target bearing under the plurality of preset diagnosis models according to the weight of each preset diagnosis model in the plurality of preset diagnosis models, to obtain the fault diagnosis result of the target bearing.

[0008] Optionally, the determining of the fault data proportion threshold of each preset diagnosis model according to the statistical value of the fault data proportion corresponding to each preset diagnosis model comprises: obtaining a preset reference threshold; and taking a larger value between the preset reference threshold and the statistical value of the fault data proportion corresponding to each preset diagnosis model as the fault data proportion threshold of each preset diagnosis model.

[0009] Optionally, the obtaining of the weight of each preset diagnosis model in the plurality of preset diagnosis models comprises: obtaining a sample vibration data set, wherein the sample vibration data set comprises a plurality of pairs of sample vibration data and fault labels, and the fault labels indicate faults or normality; processing the sample vibration data by each preset diagnosis model in the plurality of preset diagnosis models to determine whether the sample vibration data is fault data as a diagnosis result; determining an accuracy rate and a recall rate of each preset diagnosis model according to the diagnosis result of each preset diagnosis model and the fault label; and calculating a sum value of the accuracy rate and the recall rate of each preset diagnosis model as the weight of each preset diagnosis model.

[0010] Optionally, the plurality of preset diagnosis models comprises a rule diagnosis model, and the processing of each of the plurality of preset diagnosis models on the original vibration data set to determine whether each of the original vibration data in the original vibration data set is fault data comprises: performing time-frequency domain conversion processing on each of the original vibration data in the original vibration data set to obtain a plurality of pairs of vibration amplitudes and vibration frequencies corresponding to each of the original vibration data; extracting a plurality of vibration peak values from the plurality of vibration amplitudes corresponding to each of the original vibration data by the rule diagnosis model, and recording vibration frequencies corresponding to the plurality of vibration peak values as peak frequencies; obtaining a bearing fault frequency; for each of the original vibration data, in a case where there is a peak frequency satisfying a preset condition among the plurality of peak frequencies corresponding to the original vibration data, determining that the original vibration data is fault data, wherein the preset condition comprises: in frequency difference values formed by the peak frequency and other peak frequencies, there are at least N frequency difference values respectively close to at least N times of the bearing fault frequency, N is a preset positive integer, the at least N times of the bearing fault frequency comprises 1 time, and the preset condition further comprises: the frequency difference value formed by the peak frequency close to 1 time of the bearing fault frequency has the highest frequency in all frequency difference values that can be formed by the plurality of peak frequencies corresponding to the original vibration data.

[0011] Optionally, the at least N times of the bearing fault frequency comprises: a difference value of at least N times of the bearing fault frequency is less than or equal to a set multiple of a spectrum resolution, the spectrum resolution is a difference value of adjacent two spectrum lines in a frequency spectrum obtained after the time-frequency domain conversion processing, and the set multiple is a positive integer.

[0012] Optionally, the determining, for each original vibration data, that the original vibration data is fault data in a case where there is a peak frequency satisfying a preset condition among the peak frequencies corresponding to the original vibration data, comprises: determining, for each peak frequency among the peak frequencies corresponding to each original vibration data, a first frequency difference between the peak frequency and other peak frequencies among the peak frequencies corresponding to the original vibration data, and saving the first frequency difference in association with the peak frequency in a case where the first frequency difference is close to 1 times the bearing fault frequency, to obtain a plurality of first frequency differences corresponding to the original vibration data, wherein each first frequency difference in the plurality of first frequency differences is associated with at least one peak frequency; determining, from the plurality of first frequency differences corresponding to each original vibration data, a first frequency difference with a largest number of associated peak frequencies as a candidate fault frequency corresponding to each original vibration data; and determining, for each original vibration data, that the original vibration data is fault data in a case where there is at least one peak frequency satisfying a preset sub-condition among all peak frequencies associated with the candidate fault frequency corresponding to the original vibration data, wherein the preset sub-condition comprises: there are at least N second frequency differences close to at least N times the bearing fault frequency among second frequency differences formed by the peak frequency and other peak frequencies among the peak frequencies corresponding to the original vibration data.

[0013] Optionally, the plurality of preset diagnosis models comprise a shallow machine learning diagnosis model based on feature engineering, wherein features used by the shallow machine learning diagnosis model based on feature engineering comprise at least one of a main frequency band position change and a frequency spectrum dispersion degree, wherein the main frequency band position change represents a degree to which main vibration frequencies of original vibration data are concentrated in high frequencies or low frequencies, the main vibration frequencies being vibration frequencies with relatively large vibration amplitudes among a plurality of vibration frequencies obtained by time-frequency conversion processing of the original vibration data, the high frequencies or the low frequencies being determined based on an analysis frequency of the original vibration data, and the frequency spectrum dispersion degree represents a degree to which the plurality of vibration frequencies obtained by the time-frequency conversion processing of the original vibration data are concentrated or dispersed with respect to the main vibration frequencies.

[0014] Optionally, the plurality of preset diagnosis models comprise a deep learning model.

[0015] In another general aspect, a bearing fault diagnosis apparatus is provided, including: an acquisition unit configured to acquire a set of original vibration data of a target bearing; a diagnosis unit configured to process the set of original vibration data by each of a plurality of preset diagnosis models to determine whether each of the set of original vibration data is fault data, and to calculate a fault data proportion of the set of original vibration data, wherein each preset diagnosis model corresponds to a fault data proportion; the acquisition unit is further configured to acquire a weight of each of the plurality of preset diagnosis models; and a fusion unit configured to determine a fault diagnosis result of the target bearing according to the weight and the corresponding fault data proportion of each of the plurality of preset diagnosis models.

[0016] Optionally, the number of target bearings is at least one, wherein the diagnosis unit is further configured to process the set of original vibration data of each of the target bearings by each of the plurality of preset diagnosis models to determine whether each of the set of original vibration data of each of the target bearings is fault data, and to calculate a fault data proportion of the set of original vibration data of each of the target bearings.

[0017] Optionally, the bearing fault diagnosis apparatus further includes a statistics unit configured to: for each preset diagnosis model, determine a statistical value of the fault data proportion of the set of original vibration data of each of the target bearings; and determine a fault data proportion threshold of each preset diagnosis model according to the statistical value of the corresponding fault data proportion of each preset diagnosis model; wherein the fusion unit is further configured to determine the fault diagnosis result of the target bearing according to the weight, the corresponding fault data proportion, and the fault data proportion threshold of each of the plurality of preset diagnosis models.

[0018] Optionally, the fusion unit is further configured to: for each preset diagnosis model, determine a ratio of the fault data proportion of the target bearing to the fault data proportion threshold to obtain a preliminary diagnosis result of the target bearing under each preset diagnosis model; and determine a weighted average of the preliminary diagnosis results of the target bearing under the plurality of preset diagnosis models according to the weight of each of the plurality of preset diagnosis models to obtain the fault diagnosis result of the target bearing.

[0019] Optionally, the statistics unit is further configured to: acquire a preset reference threshold; and take a larger value between the preset reference threshold and the statistical value of the corresponding fault data proportion of each preset diagnosis model as the fault data proportion threshold of each preset diagnosis model.

[0020] Optionally, the obtaining unit is further configured to: obtain a sample vibration data set, wherein the sample vibration data set comprises a plurality of pairs of sample vibration data and fault labels, and the fault labels indicate faults or normality; process the sample vibration data by each of the plurality of preset diagnosis models to determine whether the sample vibration data is fault data as a diagnosis result; determine an accuracy rate and a recall rate of each preset diagnosis model according to the diagnosis result of each preset diagnosis model and the fault labels; and calculate a sum value of the accuracy rate and the recall rate of each preset diagnosis model as a weight of each preset diagnosis model.

[0021] Optionally, the plurality of preset diagnosis models comprises a rule diagnosis model, and the diagnosis unit is further configured to: perform time-frequency domain conversion processing on each of the original vibration data in the original vibration data set to obtain a plurality of pairs of vibration amplitudes and vibration frequencies corresponding to each original vibration data; extract a plurality of vibration peak values from the plurality of vibration amplitudes corresponding to each original vibration data by the rule diagnosis model, and record vibration frequencies corresponding to the plurality of vibration peak values as peak frequencies; obtain a bearing fault frequency; and for each original vibration data, in a case where there is a peak frequency satisfying a preset condition among the plurality of peak frequencies corresponding to the original vibration data, determine that the original vibration data is fault data, wherein the preset condition comprises: among frequency difference values formed by the peak frequency and other peak frequencies, there are at least N frequency difference values respectively close to at least N multiple frequencies of the bearing fault frequency, N is a preset positive integer, and the at least N multiple frequencies include 1 multiple frequency, and the preset condition further comprises: a frequency difference value formed by the peak frequency close to the 1 multiple frequency of the bearing fault frequency has the highest frequency in all frequency difference values that can be formed by the plurality of peak frequencies corresponding to the original vibration data.

[0022] Optionally, the at least N multiple frequencies close to the bearing fault frequency comprise: a difference value of the at least N multiple frequencies of the bearing fault frequency is less than or equal to a set multiple of a spectral resolution, the spectral resolution is a difference value between two adjacent spectral lines in a frequency spectrum obtained after time-frequency domain conversion processing, and the set multiple is a positive integer.

[0023] Optionally, the diagnosis unit is further configured to: for each peak frequency in the plurality of peak frequencies corresponding to each raw vibration data, determine a first frequency difference between the peak frequency and other peak frequencies in the plurality of peak frequencies corresponding to the raw vibration data, and save the first frequency difference in association with the peak frequency if the first frequency difference is close to a 1st-order frequency of the bearing fault frequency, to obtain a plurality of first frequency differences corresponding to the raw vibration data, wherein each first frequency difference in the plurality of first frequency differences is associated with at least one peak frequency; determine, from the plurality of first frequency differences corresponding to each raw vibration data, a first frequency difference with a largest number of associated peak frequencies as a candidate fault frequency corresponding to each raw vibration data; and determine the raw vibration data as fault data if there is at least one peak frequency satisfying a preset sub-condition among all peak frequencies associated with the candidate fault frequency corresponding to the raw vibration data, wherein the preset sub-condition comprises that there are at least N second frequency differences close to at least N orders of the bearing fault frequency among second frequency differences formed by the peak frequency and other peak frequencies in the plurality of peak frequencies corresponding to the raw vibration data.

[0024] Optionally, the plurality of preset diagnosis models comprises a shallow machine learning diagnosis model based on feature engineering, wherein features used by the shallow machine learning diagnosis model based on feature engineering comprise at least one of a main frequency band position change and a frequency spectrum dispersion degree, wherein the main frequency band position change represents a degree of concentration of main vibration frequencies of raw vibration data at high frequencies or low frequencies, the main vibration frequencies being vibration frequencies with relatively large amplitudes in a plurality of vibration frequencies obtained by time-frequency conversion processing of the raw vibration data, the high frequencies or the low frequencies being determined based on an analysis frequency of the raw vibration data, and the frequency spectrum dispersion degree represents a degree of concentration or dispersion distribution of a plurality of vibration frequencies obtained by time-frequency conversion processing of the raw vibration data relative to the main vibration frequencies.

[0025] Optionally, the plurality of preset diagnosis models comprises a deep learning model.

[0026] In another general aspect, a computer-readable storage medium is provided that, when instructions in the computer-readable storage medium are executed by at least one processor, cause the at least one processor to perform the bearing fault diagnosis method as described above.

[0027] In another general aspect, a computer device is provided that includes at least one processor and at least one memory storing computer-executable instructions that, when executed by the at least one processor, cause the at least one processor to perform the bearing fault diagnosis method as described above.

[0028] The present disclosure provides a bearing fault diagnosis method, device, storage medium and computer equipment. By calculating the fault data proportion obtained after diagnosis by different preset diagnosis models, the diagnosis result difference of different preset diagnosis models can be effectively reflected by one data. On this basis, by combining model weights to fuse these fault data proportions, the comprehensive fault diagnosis result is reflected by the fusion result, which can weaken the effect difference of different models in the face of different bearings and different working conditions, obtain more accurate diagnosis results compared with single model, and thus improve the precision and recall rate, help to improve the generalization of the method, reduce the false positive rate and false negative rate, and ensure the reliable diagnosis of bearing faults.

[0029] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0030] The above and other objects and features of the present application will become more apparent from the following description of embodiments taken in conjunction with the accompanying drawings, in which:

[0031] Figure 1 is a flow chart illustrating a bearing fault diagnosis method according to an embodiment of the present disclosure;

[0032] Figure 2 is a flow chart illustrating a bearing fault diagnosis method according to one specific embodiment of the present disclosure;

[0033] Figure 3 is a diagnosis flow chart illustrating a rule diagnosis model according to one specific embodiment of the present disclosure;

[0034] Figure 4 is a schematic diagram of a part of a structure of a wind turbine generator set according to one diagnosis example of the present disclosure;

[0035] Figure 5 is a block diagram of a bearing fault diagnosis device according to an embodiment of the present disclosure;

[0036] Figure 6 is a block diagram of a computer equipment according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0037] The following detailed description is presented to aid the reader in gaining a comprehensive understanding of the methods, apparatuses, and / or systems described herein. However, various changes, modifications, and equivalents can be used, and thus particular embodiments described herein are not intended as being exhaustive of the ways in which the methods, apparatuses, and / or systems described herein can be practiced. For instance, the order in which operations are described is not intended to be limiting, except in cases where a particular order is essential, for example, where a specific sequence is required as described in the assertions presented herein. Additionally, descriptions of features in terms of other features when provided is not intended to be limiting, such descriptions are only for use in providing an overall description of embodiments disclosed herein. Furthermore, descriptions of the features in terms of "first," "second," "third," and so on do not imply a particular ordering, but rather are used to name different features to assist the reader.

[0038] The features described herein can be implemented in different forms and should not be construed as limited to the examples described herein. Rather, these examples are provided as illustrative of only a few of the many possible implementations of the methods, apparatuses, and / or systems described herein.

[0039] As used herein, the term "and / or" includes any one of the associated listed items, as well as any combination of any two or more of the associated listed items.

[0040] Although terms such as "first," "second," and "third" can be used herein to describe various elements, components, regions, layers, or sections, these elements, components, regions, layers, or sections should not be limited by these terms. Rather, these terms are only used to distinguish one element, component, region, layer, or section from another element, component, region, layer, or section. Thus, a first element, component, region, layer, or section referred to in the examples described herein can also be called a second element, component, region, layer, or section without departing from the teachings of the examples.

[0041] In the description, when an element such as a layer, a region, or a substrate is referred to as being "on" another element, "connected to" or "coupled to" another element, it can be directly on, directly connected to, or directly coupled to the other element, or one or more other elements can be interposed therebetween. In contrast, when an element is referred to as being "directly on," "directly connected to," or "directly coupled to" another element, there are no other elements interposed therebetween.

[0042] The terminology used herein is for the purpose of describing various examples only and is not intended to be limiting of the disclosure. Singular forms are intended to include the plural forms, unless the context clearly indicates otherwise. The terms "comprises," "comprising," and "including" specify the presence of stated features, numbers, operations, components, elements, and / or combinations thereof, but do not preclude the presence or addition of one or more other features, numbers, operations, components, elements, and / or combinations thereof.

[0043] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. Terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and this disclosure and should not be interpreted in an overly idealized or formal sense.

[0044] Also, in the description of examples, when it is considered that a detailed description of the related structure or function that is well known will cause a blurred explanation of the present disclosure, such a detailed description will be omitted.

[0045] Figure 1 is a flowchart showing a bearing fault diagnosis method according to an embodiment of the present disclosure.

[0046] Referring to Figure 1 In step S101, a set of original vibration data of a target bearing is acquired.

[0047] In step S102, the set of original vibration data is processed by each of a plurality of preset diagnosis models to determine whether each of the set of original vibration data is fault data, and to calculate a fault data proportion of the set of original vibration data, wherein each preset diagnosis model corresponds to a fault data proportion. This step uses a plurality of preset diagnosis models to respectively preliminarily diagnose the set of original vibration data, and obtains the corresponding fault data proportion as the preliminary diagnosis result of the corresponding preset diagnosis model, which is used in the subsequent steps. It should be understood that the greater the fault data proportion, the higher the possibility of the target bearing having a fault based on the corresponding preset diagnosis model.

[0048] In step S103, a weight of each of the plurality of preset diagnosis models is acquired. The weight can reflect the reference value of each preset diagnosis model.

[0049] Optionally, the step S103 comprises: obtaining a sample vibration data set, wherein the sample vibration data set comprises a plurality of pairs of sample vibration data and fault labels, and the fault labels indicate faults or normality; processing the sample vibration data by each of a plurality of preset diagnosis models to determine whether the sample vibration data is fault data as a diagnosis result; determining an accuracy rate and a recall rate of each preset diagnosis model according to the diagnosis result of each preset diagnosis model and the fault labels; and calculating a sum value of the accuracy rate and the recall rate of each preset diagnosis model as a weight of each preset diagnosis model. The accuracy rate and the recall rate can effectively reflect the diagnosis effect of the preset diagnosis model from different angles. By taking the sum value of the accuracy rate and the recall rate as the weight of the corresponding model, the weight can represent the diagnosis effect of the model, and then the reference value of the preliminary diagnosis result of the corresponding model in the fusion process is determined according to the diagnosis effect, which helps to improve the accuracy and recall of the fault diagnosis result after fusion, and ensures the reliable diagnosis of the bearing fault.

[0050] In step S104, a fault diagnosis result of the target bearing is determined according to the weight of each of the plurality of preset diagnosis models and the corresponding proportion of fault data. This step fuses the initial diagnosis results of the plurality of preset diagnosis models based on the weight, for example, a weighted average value of the proportions of fault data corresponding to the plurality of preset diagnosis models can be calculated as the final fault diagnosis result.

[0051] According to the bearing fault diagnosis method of the example embodiment of the present disclosure, by calculating the proportion of fault data obtained after diagnosis by different preset diagnosis models, the diagnosis result difference of different preset diagnosis models can be effectively reflected by one data. On this basis, by combining the model weight to fuse these proportions of fault data, the comprehensive fault diagnosis result is reflected by the fusion result, which can weaken the effect difference of different models in the face of different bearings and different working conditions, and obtain more accurate diagnosis results compared with a single model, thereby improving the accuracy and recall rate, helping to improve the generalization of the method, reducing the false positive rate and the false negative rate, and ensuring the reliable diagnosis of the bearing fault.

[0052] Next, how to fuse the preliminary diagnosis results of the plurality of preset diagnosis models is specifically introduced.

[0053] Optionally, the number of target bearings is at least one, and the step S102 comprises: processing the original vibration data set of each target bearing by each of the plurality of preset diagnosis models to determine whether each original vibration data in the original vibration data set of each target bearing is fault data, and calculating the proportion of fault data of the original vibration data set of each target bearing. By designing the number of target bearings to be at least one and performing initial diagnosis on each target bearing, batch diagnosis of a plurality of target bearings can be achieved, which helps to improve the diagnosis efficiency.

[0054] Further optionally, the bearing fault diagnosis method according to the embodiments of the present disclosure further comprises: determining, for each preset diagnosis model, a statistical value (such as a mean value, a mode, a specific quantile, etc.) of the fault data proportion of the original vibration data set of the respective target bearing; determining, according to the statistical value of the fault data proportion corresponding to each preset diagnosis model, a fault data proportion threshold of each preset diagnosis model; wherein, step S104 comprises: determining the fault diagnosis result of the target bearing according to the weight, the corresponding fault data proportion and the fault data proportion threshold of each preset diagnosis model in the plurality of preset diagnosis models. Although theoretically, the fault data proportion can reflect the possibility of the existence of faults in the target bearing, strictly speaking, for different batches of model diagnosis calculations, even if the fault data proportions of two target bearings are the same, the actual probabilities of faults of the two target bearings are still different, and there is a diagnosis difference caused by the target bearing itself. By statistically obtaining the statistical value of the fault data proportion of all target bearings obtained by a preset diagnosis model in this diagnosis, the statistical value can be used as a reference value for this diagnosis, thereby reducing the error caused by the difference between different bearings in each diagnosis, and improving the reliability of the diagnosis result. It should be understood that, for a preset diagnosis model, if the fault data proportion of a target bearing under the model exceeds the corresponding fault data proportion threshold, it can be considered that the target bearing is diagnosed to have a fault under the model. As an example, the diagnosed target bearings can include a plurality of bearings in a plurality of units, at this time, in order to simplify data processing, bearing fault diagnosis can also be performed for each unit, that is, data processing is no longer performed in units of a single target bearing, but in units of a single unit, the data of a plurality of target bearings in the unit are regarded as the data of a single target bearing, the fault data proportion is calculated, and diagnosis processing is performed, which is also an implementation manner of the present disclosure.

[0055] It should also be understood that the more the number of target bearings / units in the current batch of diagnosis, the better the reference effect of the statistical value of the fault data proportion; when the number of target bearings / units is small, the reference effect of the statistical value is reduced, especially when the number of target bearings / units is 1, the fault data proportion is only 1, at this time, the statistical value of the fault data proportion is probably the fault data proportion itself. Therefore, for these cases where the statistical value reference effect is insufficient, additional data processing can be performed to improve the reference effect of the fault data proportion threshold.

[0056] As an example, a preset reference threshold value large enough can be configured, and accordingly, the step of determining the fault data proportion threshold value of each preset diagnostic model according to the statistical value of the fault data proportion corresponding to each preset diagnostic model includes: obtaining the preset reference threshold value; and taking the larger value between the preset reference threshold value and the statistical value of the fault data proportion corresponding to each preset diagnostic model as the fault data proportion threshold value of each preset diagnostic model. By increasing a preset reference threshold value large enough and taking the larger value between the preset reference threshold value and the statistical value as the fault data proportion threshold value, it can be ensured that the fault data proportion threshold value used is always greater than or equal to the preset reference threshold value, thereby providing a threshold lower limit, compensating for the case of insufficient statistical data, ensuring the reference effect of the obtained fault data proportion threshold value, and reducing the risk of misdiagnosis of the target bearing. It should be understood that the preset reference threshold value can be obtained by experience or through experiments, and the present disclosure does not limit this.

[0057] Optionally, the step of determining the fault diagnosis result of the target bearing according to the weight, the corresponding fault data proportion and the fault data proportion threshold value of each preset diagnostic model in the plurality of preset diagnostic models includes: for each preset diagnostic model, determining the ratio of the fault data proportion of the target bearing to the fault data proportion threshold value, to obtain a preliminary diagnosis result of the target bearing under each preset diagnostic model; and determining the weighted average value of the preliminary diagnosis results of the target bearing under the plurality of preset diagnostic models according to the weight of each preset diagnostic model in the plurality of preset diagnostic models, to obtain the fault diagnosis result of the target bearing. As to how to determine the fault diagnosis result of the target bearing in combination with the fault data proportion threshold value of each preset diagnostic model, by first calculating the ratio of the fault data proportion to the fault data proportion threshold value of the corresponding preset diagnostic model as the preliminary diagnosis result of the model, it is equivalent to using the fault data proportion threshold value to carry out non-dimensionalization processing on the fault data proportion, so that when the preliminary diagnosis result is greater than or equal to 1, it is considered that the target bearing is diagnosed to have a fault under the model. And at this time, after calculating the weighted average value of the preliminary diagnosis results of the plurality of preset diagnostic models, the final fault diagnosis result of the target bearing can still be determined based on the size relationship between the weighted average value and 1 based on this principle, without other processing, so that convenient fault diagnosis is realized by using simple data processing.

[0058] Next, different preset diagnostic models will be introduced respectively.

[0059] In some embodiments, optionally, the plurality of preset diagnosis models comprises a rule diagnosis model, and correspondingly, the step of processing the original vibration data set by each of the plurality of preset diagnosis models in step S102 to determine whether each of the original vibration data in the original vibration data set is fault data comprises: performing time-frequency domain conversion processing on each of the original vibration data in the original vibration data set to obtain a plurality of pairs of vibration amplitudes and vibration frequencies corresponding to each of the original vibration data; extracting a plurality of vibration peak values from the plurality of vibration amplitudes corresponding to each of the original vibration data by the rule diagnosis model, and recording the vibration frequencies corresponding to the plurality of vibration peak values as peak frequencies; obtaining a bearing fault frequency; for each of the original vibration data, in the plurality of peak frequencies corresponding to the original vibration data, if there is a peak frequency satisfying a preset condition, it is determined that the original vibration data is fault data, wherein the preset condition comprises: in the frequency difference values formed by the peak frequency and other peak frequencies, there are at least N frequency difference values respectively close to at least N times of the bearing fault frequency, N is a preset positive integer, and the at least N times of the bearing fault frequency includes 1 time, and the preset condition further comprises: the frequency difference value formed by the peak frequency close to 1 time of the bearing fault frequency has the highest frequency in all frequency difference values that can be formed by the plurality of peak frequencies corresponding to the original vibration data. By using the rule diagnosis model based on peak value extraction and frequency recognition, the preset condition satisfied by the fault data can be constructed by using the physical characteristics at the time of fault, and then the fault diagnosis can be realized, which can reduce the data calculation amount and ensure that the diagnosis result has a physical theory basis, and subsequent optimization diagnosis can be realized by optimizing the preset condition. Specifically, the bearing fault frequency is calculated according to the insulation diameter, the inclination angle and the number of rollers of the bearing, which is used to reflect the theoretical frequency when a specific part of the bearing fails. In other words, for different parts of the fault, there are different bearing fault frequencies, so the relationship between the vibration data of the bearing and the frequency of the bearing fault frequency of a certain part can be detected to diagnose whether the part fails. However, due to the influence of many factors during the actual operation of the bearing, the actual frequency at the time of failure deviates from the theoretically calculated bearing fault frequency. The frequency difference value between two peak frequencies can reflect the vibration situation. If the frequency difference value of a certain peak frequency is close to multiple times of the bearing fault frequency (including 1 time), and the frequency difference value formed by the peak frequency is close to the bearing fault frequency multiple times, then the frequency difference value formed by the peak frequency close to the bearing fault frequency is considered to be the actual fault frequency, and the bearing is determined to fail accordingly. As an example, N=3. It should be noted that the frequency domain data (i.e. frequency spectrum, containing a plurality of pairs of vibration amplitudes and vibration frequencies) is usually used for vibration analysis, while the data detected by the sensor is usually time domain data, so the detected time domain data needs to be converted to frequency domain data by time-frequency domain conversion processing, and then the frequency domain data is input into the model for diagnosis.The conversion processing steps here can be regarded as data preparation before formal preliminary diagnosis using various preset diagnostic models.

[0060] Further optionally, the at least N times of the bearing fault frequency close to the bearing fault frequency include: the difference value with the at least N times of the bearing fault frequency is less than or equal to a set multiple of the spectral resolution, the spectral resolution is the difference value between two adjacent spectral lines in the frequency spectrum obtained after the time-frequency conversion processing, and the set multiple is a positive integer. By explicitly defining "close to" as the difference value being less than or equal to the set multiple of the spectral resolution, a clear basis for judgment can be provided, which is helpful for realizing automatic diagnosis of the computer. In addition, the spectral resolution reflects the accuracy of the time-frequency conversion processing, and by specifically taking the set multiple of the spectral resolution as the reference amount for judgment, the judgment can be adapted to the actual time-frequency conversion processing, thereby providing a more reasonable basis for judgment when facing different time-frequency conversion processes. It should be understood that the set multiple can be obtained through experiments or according to experience, and the present disclosure does not limit this. As an example, the set multiple is 2.

[0061] Further optionally, for each original vibration data, when there is a peak frequency that meets a preset condition among the multiple peak frequencies corresponding to the original vibration data, the original vibration data is determined to be fault data, including: for each peak frequency among the multiple peak frequencies corresponding to each original vibration data, determining a first frequency difference between the peak frequency and other peak frequencies among the multiple peak frequencies corresponding to the original vibration data, and when the first frequency difference is close to 1 times the frequency of the bearing fault frequency, the first frequency difference is associated with the peak frequency and saved to obtain multiple first frequency differences corresponding to the original vibration data, wherein each of the multiple first frequency differences is associated with at least one Peak frequency; from the multiple first frequency difference values ​​corresponding to each original vibration data, determine the first frequency difference value with the largest number of associated peak frequencies as the candidate fault frequency corresponding to each original vibration data; for each original vibration data, if there is at least one peak frequency that meets the preset sub-condition among all the peak frequencies associated with the candidate fault frequency corresponding to the original vibration data, determine that the original vibration data is fault data, wherein the preset sub-condition includes: in the second frequency difference value formed by the peak frequency and other peak frequencies among the multiple peak frequencies corresponding to the original vibration data, there are at least N second frequency differences, each of which is close to at least N multiples of the bearing fault frequency. The first two steps can ensure that the candidate fault frequency obtained thereby has the highest frequency of occurrence among all the frequency differences that can be formed by the multiple peak frequencies corresponding to the original vibration data, which is equivalent to meeting the second condition in the preset conditions. At the same time, since the candidate fault frequency is close to 1 multiple of the bearing fault frequency, each peak frequency associated with the candidate fault frequency can meet the 1 multiple condition in the first condition of the preset conditions. On this basis, the third step can be directly executed for all peak frequencies associated with the candidate fault frequency to determine whether the candidate fault frequency meets the first of the preset conditions, without having to determine whether each peak frequency meets the 1-time frequency condition. This judgment sequence effectively reduces the amount of calculation and improves diagnostic efficiency while meeting the judgment requirements.

[0062] As an example, specifically, after peak extraction, a peak list can be obtained, which contains multiple peak frequencies. First, traverse the peak list, for the current peak frequency A i , calculate the frequencies B of all other peaks in the peak list j (j≠i) and A i The difference between j -A i - bearing fault frequency)≤2 times the spectrum resolution, it can be considered that (B j -A i ) is close to the bearing failure frequency, and the (B j -A i ) and Ai The (B-A) and the corresponding A are obtained through traversal, and since (B-A) is a difference, the same (B-A) can correspond to multiple As. The (B-A) corresponding to the maximum A is taken as the candidate fault frequency. Then, multiple As corresponding to the candidate fault frequency are traversed in a similar manner, except that the (B-A) is judged to be close to 1x frequency before, and here, the (B-A) is judged to be close to multiple frequencies less than the highest frequency, respectively. For example, if the highest frequency obtained after the time-frequency conversion processing is between 4x frequency and 5x frequency, then 1x frequency to 4x frequency is judged here. If at least one A of the multiple As traversed can satisfy the judgment condition of 3x frequency or more, for example, there is an A, and there are 3 (B-A) close to 1x frequency, 2x frequency and 4x frequency, respectively, then the candidate fault frequency is considered to be the actual fault frequency, and the target bearing fault is determined. j -A i ) whether it is close to 1x frequency, and here, the (B-A) is judged to be close to multiple frequencies less than the highest frequency, respectively. For example, if the highest frequency obtained after the time-frequency conversion processing is between 4x frequency and 5x frequency, then 1x frequency to 4x frequency is judged here. If at least one A of the multiple As traversed can satisfy the judgment condition of 3x frequency or more, for example, there is an A, and there are 3 (B-A) close to 1x frequency, 2x frequency and 4x frequency, respectively, then the candidate fault frequency is considered to be the actual fault frequency, and the target bearing fault is determined.

[0063] In some embodiments, optionally, the plurality of preset diagnosis models comprises a shallow machine learning diagnosis model based on feature engineering, wherein the features used by the shallow machine learning diagnosis model based on feature engineering comprise at least one of the following: main frequency band position change, frequency spectrum dispersion degree, wherein the main frequency band position change represents the degree to which the main vibration frequency of the original vibration data is concentrated in high or low frequency, the main vibration frequency being a vibration frequency with relatively large amplitude in the plurality of vibration frequencies obtained by the original vibration data through time-frequency conversion processing, the high or low frequency being determined based on the analysis frequency of the original vibration data, and the frequency spectrum dispersion degree representing the degree to which the plurality of vibration frequencies obtained by the original vibration data through time-frequency conversion processing are distributed in a concentrated or dispersed manner with respect to the main vibration frequency. By adopting the shallow machine learning diagnosis model based on feature engineering and introducing at least one of the main frequency band position change and the frequency spectrum dispersion degree as the features used by the model, the diagnosis effect of the model can be effectively improved. Specifically, the main frequency band position change reflects the component characteristics of the main vibration frequency, and the frequency spectrum dispersion degree reflects the distribution relationship between other vibration frequencies and the main vibration frequency, both of which can reflect the frequency characteristics of the original vibration data from different aspects. It should be understood that, for the main frequency band position change, the high frequency and the low frequency refer to the relatively high and relatively low frequencies in the analysis frequency range of the vibration data, the minimum analysis frequency is 0, and the maximum analysis frequency is the value obtained by dividing the sampling frequency by 2.56 (2 is used in textbooks and 2.56 is used in the industry), for example, when the sampling frequency is 1280 Hz, the maximum analysis frequency is 500 Hz, so the analysis frequency is related to the sampling frequency, and data exceeding the maximum value is greatly affected by other factors and has low accuracy, so it is not analyzed. Accordingly, low-pass filtering is performed during time-frequency conversion processing to filter out components with extremely high frequencies.

[0064] As an example, the main frequency band position change is calculated by the following formula:

[0065]

[0066] wherein s k is the frequency amplitude sequence of the original vibration data, k = 1, 2, 3, …, K is the number of spectral lines; f k is the frequency value of the kth spectral line. The greater the value of fe1, the more concentrated the main vibration frequency is in high frequency, and vice versa.

[0067] The spectral dispersion degree is calculated by the following formula:

[0068]

[0069] The meanings of the symbols are as described above. The smaller the value of fe2, the more vibration frequencies around the main vibration frequency in the data, and the more concentrated the spectrum, and vice versa.

[0070] It should be understood that in addition to the main frequency band position change and the spectral dispersion degree, other conventional features can also be used in the shallow machine learning diagnosis model based on feature engineering to utilize sufficient information to ensure reliable diagnosis.

[0071] Optionally, the plurality of preset diagnosis models includes a deep learning model. The deep learning model has a deep structure and strong non-linear feature extraction capability, and can directly realize extraction of fault features and fault pattern recognition in the original vibration data, which is conducive to realizing end-to-end fault diagnosis under complex working conditions. By using the deep learning model, the selection range of the preset diagnosis model can be widened, which is helpful to improve the final diagnosis effect.

[0072] Next, a bearing fault diagnosis method according to one specific embodiment of the present disclosure is introduced. Figures 2 to 5

[0073] ​In this specific embodiment, the preset diagnosis models used include a mechanism-based rule diagnosis model, an XGBoost diagnosis model based on feature engineering, and a CNN (Convolutional Neural Network) diagnosis model based on deep learning. First, the XGBoost diagnosis model and the CNN diagnosis model are pre-trained based on a historical fault case library. The historical fault case library has multiple historical labeled data, that is, historical vibration data with labels. The labels indicate that the corresponding bearing of the data is faulty or normal. Each data can be directly stored in the form of a frequency spectrum to reduce the data storage amount, and the frequency spectrum can be directly used when the model is pre-trained. The trained model is used as a preset diagnosis model. These preset diagnosis models do not update parameters or train during diagnosis, but only make judgments on data. For fault diagnosis of multiple target bearings, the data in the original vibration data set uses transmission chain vibration monitoring data, and the time domain data detected can be uniformly processed for time domain and frequency domain conversion, for example, using an FFT (fast Fourier transform) method for processing to obtain frequency domain data. Each preset diagnosis model takes the frequency domain data as input. The process is shown in FIG. 1. Figure 2 .

[0074] Specifically, first, the XGBoost diagnosis model and the CNN diagnosis model are pre-trained. The specific training method belongs to mature technology in the art, and will not be described here. After pre-training is completed, the sum of the precision and recall of each preset diagnosis model is calculated as the weight of the model.

[0075] Then, before formal diagnosis, the original vibration data to be diagnosed can be subjected to data quality detection, that is, data cleaning. The difference between data cleaning here and conventional data cleaning is that, for this special form of movement, vibration, abnormal data that does not repeatedly change around 0 needs to be removed. Data cleaning mainly involves some characteristic index calculations, including mean deviation (used to detect whether the data is offset near 0), ratio of the number of data points with the same value to the total data volume (not too many), ratio of the difference between the number of positive and negative data points to the total number of data points (reflecting whether the data changes uniformly around 0), positive impact of vibration data (indicating data that is too large and exceeds the normal vibration limit, for example, if the amplitude of a certain data is 10 or more, it needs to be removed), drift anomaly threshold (zero-crossing rate, similar to positive impact of vibration data), interruption of vibration data (indicating a sudden cliff-like change in data size), sliding window mean difference (if the difference between the mean of the values in the first sliding window and the mean of the values in the last sliding window is large, it indicates that the data has changed greatly and does not meet the characteristics of vibration data, for example, the mean of the first 500 data is -0.5, the mean of the last 500 data is 0.5, and the difference is -1, which is considered abnormal), and also includes effective value, kurtosis value, null value and non-real value ratio of vibration data, etc. Only data that passes the data quality detection will enter the diagnosis stage, otherwise it will be put into the database as part of the original vibration data and participate in the subsequent result statistics.

[0076] Next, preliminary diagnosis is performed using each preset diagnosis model.

[0077] The diagnosis process of the rule-based diagnosis model refers to Figure 3 , which adopts a diagnosis method based on peak extraction and frequency identification. In simple terms, the diagnosis logic is to extract peaks from the frequency spectrum, search for a frequency difference in the peak difference list formed by the corresponding peak list that is less than 2 times the spectral resolution, and when the number of frequency differences is greater than or equal to 3, it is considered to be a fault, and specifically the fault of the part corresponding to the searched bearing fault frequency. As shown in Figure 3 , the peak extraction specifically extracts two types of data, one is the largest value in the top 0.2% of all vibration frequencies in the frequency spectrum, and the other is the value greater than the sum of the mean and variance, where the mean and variance refer to the mean and variance of all vibration frequencies in the frequency spectrum.

[0078] Regarding the XGBoost diagnosis model, before the data is input into the trained XGBoost diagnosis model, the data spectrum needs to be processed, the length is reduced by downsampling, and the length is supplemented by zero. The processed spectrum meets the input of 8192 length and the highest analysis frequency of 500 Hz. Calculate the selected characteristic values of the processed spectrum, and then input the characteristic values into the XGBoost diagnosis model. The model result output is 0 or 1, 0 represents normal, and 1 represents fault.

[0079] Regarding the CNN diagnosis model, the data spectrum also needs to be processed in the aforementioned manner before being input into the trained CNN diagnosis model, and the processed spectrum is input into the CNN diagnosis model. The model result output is also 0 or 1.

[0080] Subsequently, the output values of each model are counted, and a statistical model can be used. In the preliminary diagnosis process, the statistical model records the number of data that meet the quality requirements in units of machine groups, calculates the proportion of fault data of each machine group diagnosed by different preset diagnosis models, and takes max(0.36, (avg+3*std)) as the fault data proportion threshold. Among them, avg is the average of the fault data proportions of all machine groups, std is the standard deviation of the fault data proportions of all machine groups, (avg+3*std) represents the statistical value of the fault data proportion, and 0.36 is a preset reference threshold.

[0081] Finally, the diagnosis results of each preset diagnosis model are summarized.

[0082] Suppose the fault data proportion of model i for a certain target bearing is prop i , the fault data proportion threshold is thre i , and the weight is w i , then the final model result is:

[0083]

[0084] When result is greater than or equal to 1, it is determined that the target bearing is faulty.

[0085] In one diagnosis example, based on the bearing fault diagnosis method of the above specific embodiments, the disclosure diagnoses the bearings of a batch of machine groups. The original vibration data set contains vibration acceleration data of 744 GW2X machine groups of 51 wind farms during 2022-2023, which includes 736 positive examples (normal machine groups) and 8 negative examples (problem machine groups). IEPE (Integrated Electronics Piezo-Electric) acceleration sensors are installed near the main bearing and the stator. The acceleration sensor at the main bearing position is a low-frequency acceleration sensor with a frequency range of 0.1Hz-10000Hz, and the acceleration sensor at the stator position is a full-frequency acceleration sensor with a frequency range of 0.5Hz-10000Hz. There are a total of 7 acceleration sensors, which are respectively arranged at the front bearing 41 (the front bearing 41 is located Figure 4 Figure 4 ​inside the fairing 42 in the figure, thus the dashed line indicates its position), the rear bearing 43, and the stator 44. Specifically, the front bearing 41 is provided with two sensors for detecting radial acceleration, which are respectively arranged at a horizontal position (3 o'clock position or 9 o'clock position) and a vertical position (12 o'clock position or 6 o'clock position) to detect radial acceleration in the horizontal direction and the vertical direction; the rear bearing 43 is also provided with one sensor for detecting radial acceleration at a horizontal position and a vertical position; and the stator 44 is provided with two sensors for detecting radial acceleration and one sensor for detecting axial acceleration, the former are respectively arranged at two horizontal positions, i.e., one at the 3 o'clock position and one at the 9 o'clock position, and the latter can be arranged at the edge, for example, also at the 3 o'clock position. The acceleration data of the main bearing measuring points are mainly used in this example. The data contains two kinds of collected vibration acceleration waveforms, and the collection definition 1 is that the sampling point number is 65536, the sampling frequency is 25600 Hz, the sampling period is 2.56 s, and the collection definition 2 is that the sampling point number is 65536, the sampling frequency is 1280 Hz, and the sampling period is 51.2 s.

[0086] The features screened by the feature engineering are as follows:

[0087] freq_spectrum_80_180_c3: Grouped entropy of 80-180 Hz;

[0088] freq_spectrum_20_120_index_mass_quantile: 20-120 Hz quantile index;

[0089] freq_spectrum_number_cwt_peaks_I: Peak number;

[0090] freq_spectrum_longest_strike_above_mean: Longest continuous sequence length above mean;

[0091] freq_spectrum_energy_ratio_by_chunks_8: Chunked local entropy ratio;

[0092] freq_spectrum_340_440_var: 340-440 Hz variance;

[0093] freq_spectrum_8: Main frequency band position change;

[0094] freq_spectrum_0_100_energy_ratio_by_chunks_0: chunked local entropy ratio of 0Hz-100Hz;

[0095] freq_spectrum_12: frequency spectrum spread;

[0096] freq_spectrum_0_100_agg_linear_trend: linear regression of 0Hz-100Hz based on chunked time series aggregated values;

[0097] freq_spectrum_first_location_of_minimum: minimum value location;

[0098] freq_spectrum_longest_strike_below_mean: longest consecutive strike length below mean.

[0099] The test results are shown in the following table.

[0100] Model False positives False negatives True faults Precision Recall Rule diagnostic model 10 0 8 44.44％ 100.00％ CNN diagnostic model 10 2 8 37.50％ 75.00％ XGBoost diagnostic model 13 1 8 35.00％ 87.50％ Fusion model 3 0 8 72.73％ 100.00％

[0101] In an actual fan operation scene, a wind turbine main shaft bearing fault diagnosis model is built to timely warn before a large component fails and accurately analyze the fault cause after the failure occurs, so that a reasonable maintenance plan and spare parts allocation plan can be developed, which can greatly reduce operation and maintenance costs and improve power generation. At present, single models are mainly used in actual application scenarios, and physical models, statistical models, and knowledge models each have their own advantages, but a single type cannot avoid the limitations of the model. For example, in the test results, 10 of the 744 wind turbines were misdiagnosed by the rule-based diagnosis model, 10 were misdiagnosed by the CNN diagnosis model, 2 were missed, 13 were misdiagnosed by the XGBoost diagnosis model, and 1 was missed. The present disclosure proposes a fusion scheme for multiple model results, which weights the results of multiple preset diagnosis models. For example, in the test results, the fusion model is a weighted fusion of the results of the rule-based diagnosis model, the CNN diagnosis model, and the XGBoost diagnosis model, which reduces the number of misdiagnosed units to 3 and the number of missed units to 0. Compared with the three preset diagnosis models, the fusion model has significantly improved precision and recall. The above examples prove that the fusion scheme can improve the diagnosis effect of the model, thereby more accurately diagnosing the state of the wind turbine main shaft bearing and providing more effective support for planned operation and maintenance.

[0102] Figure 5 is a block diagram illustrating a bearing fault diagnosis apparatus according to one embodiment of the present disclosure.

[0103] Referring to Figure 5The bearing fault diagnosis apparatus 500 comprises an acquisition unit 501, a diagnosis unit 502, and a fusion unit 503.

[0104] The acquisition unit 501 can acquire a set of original vibration data of a target bearing.

[0105] The diagnosis unit 502 can process the set of original vibration data through each of a plurality of preset diagnosis models to determine whether each of the set of original vibration data is fault data, and calculate a fault data proportion of the set of original vibration data, wherein each preset diagnosis model corresponds to a fault data proportion.

[0106] The acquisition unit 501 can also acquire a weight of each of the plurality of preset diagnosis models.

[0107] Optionally, the acquisition unit 501 can also acquire a set of sample vibration data, wherein the set of sample vibration data comprises a plurality of pairs of sample vibration data and fault labels, and the fault labels indicate faults or normal; process the sample vibration data through each of the plurality of preset diagnosis models to determine whether the sample vibration data is fault data as a diagnosis result; determine an accuracy and a recall rate of each preset diagnosis model according to the diagnosis result of each preset diagnosis model and the fault labels; and calculate a sum value of the accuracy and the recall rate of each preset diagnosis model as a weight of each preset diagnosis model.

[0108] The fusion unit 503 can determine a fault diagnosis result of the target bearing according to the weight and the corresponding fault data proportion of each of the plurality of preset diagnosis models.

[0109] Optionally, the number of target bearings is at least one, wherein the diagnosis unit 502 can also process the set of original vibration data of each target bearing through each of the plurality of preset diagnosis models respectively to determine whether each of the set of original vibration data of each target bearing is fault data, and calculate a fault data proportion of the set of original vibration data of each target bearing.

[0110] Optionally, the bearing fault diagnosis apparatus further comprises a statistical unit (not shown in the figure), which can: for each preset diagnosis model, determine a statistical value of the fault data proportions of the sets of original vibration data of the target bearings; and determine a fault data proportion threshold of each preset diagnosis model according to the statistical value of the fault data proportion corresponding to each preset diagnosis model; wherein the fusion unit 503 can also determine the fault diagnosis result of the target bearing according to the weight, the corresponding fault data proportion, and the fault data proportion threshold of each of the plurality of preset diagnosis models.

[0111] Optionally, the fusion unit 503 can further: for each preset diagnosis model, determine a ratio of the fault data proportion of the target bearing to the fault data proportion threshold, to obtain a preliminary diagnosis result of the target bearing under each preset diagnosis model; and determine a weighted average of the preliminary diagnosis results of the target bearing under the plurality of preset diagnosis models according to the weight of each preset diagnosis model in the plurality of preset diagnosis models, to obtain a fault diagnosis result of the target bearing.

[0112] Optionally, the statistical unit can further: obtain a preset reference threshold; and take a larger value between the preset reference threshold and the statistical value of the fault data proportion corresponding to each preset diagnosis model as the fault data proportion threshold of each preset diagnosis model.

[0113] Optionally, the plurality of preset diagnosis models include a rule diagnosis model, and the diagnosis unit 502 can further: perform time-frequency domain conversion processing on each original vibration data in the original vibration data set through the rule diagnosis model to obtain a plurality of pairs of vibration amplitudes and vibration frequencies corresponding to each original vibration data; extract a plurality of vibration peak values from the plurality of vibration amplitudes corresponding to each original vibration data, and record the vibration frequencies corresponding to the plurality of vibration peak values as peak frequencies; obtain a bearing fault frequency; and for each original vibration data, in a case where there is a peak frequency that meets a preset condition among the plurality of peak frequencies corresponding to the original vibration data, determine that the original vibration data is fault data, wherein the preset condition includes: among the frequency differences formed by the peak frequency and other peak frequencies, there are at least N frequency differences that are respectively close to at least N times of the bearing fault frequency, N is a preset positive integer, and the at least N times of the bearing fault frequency includes 1 time of the bearing fault frequency, and the preset condition further includes: the frequency difference formed by the peak frequency that is close to 1 time of the bearing fault frequency has the highest frequency among all frequency differences that can be formed by the plurality of peak frequencies corresponding to the original vibration data.

[0114] Optionally, the at least N times of the bearing fault frequency includes: a difference between the at least N times of the bearing fault frequency and the at least N times of the bearing fault frequency is less than or equal to a set multiple of a spectral resolution, the spectral resolution is a difference between two adjacent spectral lines in a frequency spectrum obtained after the time-frequency domain conversion processing, and the set multiple is a positive integer.

[0115] Optionally, the diagnosis unit 502 can further: for each peak frequency in the plurality of peak frequencies corresponding to each original vibration data, determine a first frequency difference between the peak frequency and other peak frequencies in the plurality of peak frequencies corresponding to the original vibration data, and save the first frequency difference in association with the peak frequency in a case where the first frequency difference is close to 1 times the bearing fault frequency, to obtain a plurality of first frequency differences corresponding to the original vibration data, wherein each of the plurality of first frequency differences is associated with at least one peak frequency; determine, from the plurality of first frequency differences corresponding to each original vibration data, a first frequency difference with a largest number of associated peak frequencies as a candidate fault frequency corresponding to each original vibration data; and in a case where there is at least one peak frequency satisfying a preset sub-condition among all peak frequencies associated with the candidate fault frequency corresponding to the original vibration data, determine the original vibration data as fault data, wherein the preset sub-condition comprises: there are at least N second frequency differences close to at least N times the bearing fault frequency among second frequency differences formed by the peak frequency and other peak frequencies in the plurality of peak frequencies corresponding to the original vibration data.

[0116] Optionally, the plurality of preset diagnosis models comprise a shallow machine learning diagnosis model based on feature engineering, wherein the features used by the shallow machine learning diagnosis model based on feature engineering comprise at least one of: a main frequency band position change, and a frequency spectrum dispersion degree, wherein the main frequency band position change represents a degree to which main vibration frequencies of the original vibration data are concentrated in high frequencies or low frequencies, the main vibration frequencies being vibration frequencies with relatively large vibration amplitudes in a plurality of vibration frequencies obtained by time-frequency conversion processing of the original vibration data, the high frequencies or the low frequencies being determined based on an analysis frequency of the original vibration data, and the frequency spectrum dispersion degree represents a degree to which a plurality of vibration frequencies obtained by time-frequency conversion processing of the original vibration data are concentrated or dispersed with respect to the main vibration frequencies.

[0117] Optionally, the plurality of preset diagnosis models comprise a deep learning model.

[0118] As to the apparatus in the above-described embodiments, the specific manners in which the units perform operations have been described in detail in the embodiments of the method, and thus will not be described in detail here.

[0119] The bearing fault diagnosis method according to the embodiments of the present disclosure can be written as a computer program and stored on a computer-readable storage medium. When the instructions corresponding to the computer program are executed by a processor, the bearing fault diagnosis method as described above can be implemented. Examples of the computer-readable storage medium include a read-only memory (ROM), a random access programmable read-only memory (PROM), an electrically erasable programmable read-only memory (EEPROM), a random access memory (RAM), a dynamic random access memory (DRAM), a static random access memory (SRAM), a flash memory, a non-volatile memory, a CD-ROM, a CD-R, a CD+R, a CD-RW, a CD+RW, a DVD-ROM, a DVD-R, a DVD+R, a DVD-RW, a DVD+RW, a DVD-RAM, a BD-ROM, a BD-R, a BD-RLTH, a BD-RE, a Blu-ray or an optical disc memory, a hard disk drive (HDD), a solid state drive (SSD), a card memory (such as a multimedia card, a secure digital (SD) card or an extreme digital (XD) card), a magnetic tape, a floppy disk, a magneto-optical data storage device, an optical data storage device, a hard disk, a solid state disk, and any other device configured to store a computer program and any associated data, data files and data structures in a non-transitory manner and provide the computer program and any associated data, data files and data structures to a processor or a computer so that the processor or the computer can execute the computer program. In one example, the computer program and any associated data, data files and data structures are distributed on a networked computer system, so that the computer program and any associated data, data files and data structures are stored, accessed and executed by one or more processors or computers in a distributed manner.

[0120] Figure 6 is a block diagram illustrating a computer device according to an embodiment of the present disclosure.

[0121] With reference to Figure 6 The computer device 600 includes at least one memory 601 and at least one processor 602, and the at least one memory 601 stores a set of computer executable instructions, which, when executed by the at least one processor 602, perform a bearing fault diagnosis method according to an exemplary embodiment of the present disclosure.

[0122] By way of example, computer device 600 can be a PC computer, a tablet device, a personal digital assistant, a smart phone, or other device capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken by that device. Herein, the computer device 600 need not be a single device, but can be a collection of devices or a collection of components that each have specific

[0123] In computer device 600, processor 602 can include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example, and not limitation, a processor can include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, or the like.

[0124] Processor 602 can execute instructions or code stored in memory 601, where the memory 601 can also store data. The instructions and data could alternatively be transmitted or received via a network interface device using any one of a number of well-known transfer protocols.

[0125] Memory 601 can be integrated with processor 602, such as RAM or flash memory disposed within an integrated circuit microprocessor, etc. Alternatively, memory 601 can include a separate device, such as an external disk drive, memory array or other storage device that can be used with any database system. Memory 601 and processor 602 can be operatively coupled, such as via an I / O port, a network connection, etc., so that processor 602 can read files stored in memory.

[0126] In addition, computer device 600 can further include a video display (such as a liquid crystal display) and a user input interface (such as a keyboard, a mouse, a touchscreen, etc.). All of the components of computer device 600 can be connected via a bus or other communication medium, and / or network.

[0127] The present disclosure provides a bearing fault diagnosis method and device, a storage medium and a computer device. By calculating the fault data proportion obtained after diagnosis by different preset diagnosis models, the diagnosis result difference of different preset diagnosis models can be effectively reflected by one data. On this basis, by combining model weights to fuse these fault data proportions, the comprehensive fault diagnosis result is reflected by the fusion result, which can weaken the effect difference of different models in the face of different bearings and different working conditions, obtain more accurate diagnosis results compared with single model, and thus improve the accuracy and recall rate, help to improve the generalization of the method, reduce the false positive rate and false negative rate, and ensure the reliable diagnosis of bearing faults.

[0128] The specific embodiments of the present disclosure are described in detail above, although some embodiments have been shown and described, those skilled in the art should understand that modifications and variations can be made to these embodiments without departing from the principles and spirits of the present disclosure, which are defined by the claims and their equivalents, and these modifications and variations should also be within the protection scope of the claims of the present disclosure.

Claims

1. A bearing fault diagnosis method characterized by, The bearing fault diagnosis method comprises: obtaining a set of original vibration data of target bearings, wherein the number of the target bearings is at least one; for each target bearing, processing the set of original vibration data through each of a plurality of preset diagnosis models to determine whether each original vibration data in the set of original vibration data is fault data, and calculating a fault data proportion of the set of original vibration data, wherein each preset diagnosis model corresponds to a fault data proportion; obtaining a weight of each of the plurality of preset diagnosis models; for each preset diagnosis model, determining a statistical value of the fault data proportion of the set of original vibration data of each target bearing; determining a fault data proportion threshold value of each preset diagnosis model according to the statistical value of the fault data proportion corresponding to each preset diagnosis model; for each target bearing, determining a ratio of the fault data proportion of the target bearing to the fault data proportion threshold value for each preset diagnosis model to obtain a preliminary diagnosis result of the target bearing under each preset diagnosis model, and determining a weighted average value of the preliminary diagnosis result of the target bearing under the plurality of preset diagnosis models according to the weight of each of the plurality of preset diagnosis models to obtain a fault diagnosis result of the target bearing.

2. The bearing fault diagnostic method of claim 1, wherein, The determination of the fault data proportion threshold value of each preset diagnosis model according to the statistical value of the fault data proportion corresponding to each preset diagnosis model comprises: obtaining a preset reference threshold value; taking a larger value between the preset reference threshold value and the statistical value of the fault data proportion corresponding to each preset diagnosis model as the fault data proportion threshold value of each preset diagnosis model.

3. The bearing fault diagnostic method of claim 1, wherein, The obtaining of the weight of each of the plurality of preset diagnosis models comprises: obtaining a set of sample vibration data, wherein the set of sample vibration data comprises a plurality of pairs of sample vibration data and fault labels, and the fault labels indicate faults or normality; processing the sample vibration data through each of the plurality of preset diagnosis models to determine whether the sample vibration data is fault data as a diagnosis result; determining an accuracy and a recall rate of each preset diagnosis model according to the diagnosis result of each preset diagnosis model and the fault labels; calculating a sum value of the accuracy and the recall rate of each preset diagnosis model as the weight of each preset diagnosis model.

4. The bearing fault diagnostic method of claim 1, wherein, The plurality of preset diagnosis models comprises a rule diagnosis model, wherein the processing of the set of original vibration data through each of the plurality of preset diagnosis models to determine whether each original vibration data in the set of original vibration data is fault data comprises: performing time-frequency domain conversion processing on each original vibration data in the set of original vibration data to obtain a plurality of pairs of vibration amplitudes and vibration frequencies corresponding to each original vibration data; extracting a plurality of vibration peak values from the plurality of vibration amplitudes corresponding to each original vibration data through the rule diagnosis model, and recording vibration frequencies corresponding to the plurality of vibration peak values as peak frequencies; obtaining a bearing fault frequency; In a case that, for each original vibration data, there is a peak frequency meeting a preset condition among the peak frequencies corresponding to the original vibration data, the original vibration data is determined as fault data, wherein the preset condition comprises: there are at least N frequency difference values, respectively close to at least N times of the bearing fault frequency, among the frequency difference values formed by the peak frequency and other peak frequencies, N being a preset positive integer, and the at least N times of the bearing fault frequency comprising 1 time of the bearing fault frequency, and the preset condition further comprises: the frequency difference value formed by the peak frequency, close to 1 time of the bearing fault frequency, has the highest frequency among all the frequency difference values that can be formed by the peak frequencies corresponding to the original vibration data.

5. The bearing fault diagnostic method of claim 4, wherein, The at least N times of the bearing fault frequency comprises: a difference between the at least N times of the bearing fault frequency and the peak frequency is less than or equal to a set multiple of a spectrum resolution, the spectrum resolution being a difference between two adjacent spectrum lines in a frequency spectrum obtained after time-domain frequency-domain conversion processing, and the set multiple being a positive integer.

6. The bearing fault diagnostic method of claim 4 wherein, The case that, for each original vibration data, there is a peak frequency meeting a preset condition among the peak frequencies corresponding to the original vibration data, the original vibration data is determined as fault data, comprises: For each peak frequency among the peak frequencies corresponding to each original vibration data, a first frequency difference value between the peak frequency and other peak frequencies corresponding to the original vibration data is determined, and in a case that the first frequency difference value is close to 1 time of the bearing fault frequency, the first frequency difference value is associated with the peak frequency and saved, to obtain a plurality of first frequency difference values corresponding to the original vibration data, wherein each first frequency difference value in the plurality of first frequency difference values is associated with at least one peak frequency; From the plurality of first frequency difference values corresponding to each original vibration data, a first frequency difference value associated with the most peak frequencies is determined as a candidate fault frequency corresponding to each original vibration data; In a case that, for each original vibration data, there is at least one peak frequency meeting a preset sub-condition among all the peak frequencies associated with the candidate fault frequency corresponding to the original vibration data, the original vibration data is determined as fault data, wherein the preset sub-condition comprises: there are at least N second frequency difference values, respectively close to at least N times of the bearing fault frequency, among second frequency difference values formed by the peak frequency and other peak frequencies corresponding to the original vibration data.

7. The bearing fault diagnostic method of claim 1 wherein, The plurality of preset diagnosis models comprises a shallow machine learning diagnosis model based on feature engineering, wherein the features used by the shallow machine learning diagnosis model based on feature engineering comprise at least one of a main frequency band position change and a frequency spectrum concentration degree, wherein the main frequency band position change represents a degree to which a main vibration frequency of the original vibration data is concentrated in a high frequency or a low frequency, the main vibration frequency being a vibration frequency with a relatively large amplitude among a plurality of vibration frequencies obtained by time-frequency conversion processing of the original vibration data, the high frequency or the low frequency being determined based on an analysis frequency of the original vibration data, and the frequency spectrum concentration degree represents a degree to which the plurality of vibration frequencies obtained by time-frequency conversion processing of the original vibration data are concentrated or dispersed with respect to the main vibration frequency.

8. The bearing fault diagnostic method of claim 1 wherein, The plurality of preset diagnosis models comprises a deep learning model.

9. A bearing failure diagnostic apparatus characterized by comprising: The bearing fault diagnosis apparatus comprises: An acquisition unit configured to acquire a set of original vibration data of a target bearing, wherein the number of the target bearings is at least one; A diagnosis unit configured to, for each of the target bearings, process the set of original vibration data by each of a plurality of preset diagnosis models to determine whether each of the set of original vibration data is fault data, and calculate a fault data proportion of the set of original vibration data, wherein each preset diagnosis model corresponds to a fault data proportion; The acquisition unit is further configured to acquire a weight of each of the plurality of preset diagnosis models; A statistical unit configured to, for each preset diagnosis model, determine a statistical value of the fault data proportion of the set of original vibration data of each of the target bearings, and determine a fault data proportion threshold of each preset diagnosis model according to the statistical value of the fault data proportion corresponding to each preset diagnosis model; A fusion unit configured to, for each of the target bearings, determine, for each preset diagnosis model, a ratio of the fault data proportion of the target bearing to the fault data proportion threshold to obtain a preliminary diagnosis result of the target bearing under each preset diagnosis model, and determine a weighted average of the preliminary diagnosis results of the target bearing under the plurality of preset diagnosis models according to the weight of each of the plurality of preset diagnosis models to obtain a fault diagnosis result of the target bearing.

10. A computer device, comprising: Comprise: At least one processor; At least one memory storing computer-executable instructions, Wherein the computer-executable instructions, when executed by the at least one processor, cause the at least one processor to perform the bearing fault diagnosis method of any one of claims 1 to 8.

11. A computer readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by at least one processor, cause the at least one processor to perform the bearing fault diagnosis method of any one of claims 1 to 8.

Images