Method and system for fault diagnosis through bearing noise detection

Abstract

The present disclosure provides a method and a system for fault diagnosis through bearing noise detection. The method includes: collecting noise in bearing detection, where the noise includes bearing noise and operating condition noise; preprocessing the collected noise to obtain a first time-domain signal and a second frequency-domain signal; and inputting the first time-domain signal and the second frequency-domain signal into a bearing fault diagnosis model. Among them, the bearing fault diagnosis model includes a feature extraction model and a fault discrimination model. The feature extraction model respectively extracts features from the first time-domain signal and the second frequency-domain signal to obtain a first feature associated with time-domain impact peaks and a second feature associated with fault frequency peaks. And among them, the fault discrimination model combines the first feature and the second feature and discriminates faults based on the combined features.

Description

Technical Field

[0001] The present invention relates to the field of bearing noise detection, and particularly to a method and system for fault diagnosis through bearing noise detection. Background Art

[0002] Bearing noise is one of the key parameters in bearing quality control, and certain early defects of bearings can be detected through bearing noise detection in a production line. However, noise has a strong correlation with bearing types. For example, bearings of different types have different bearing sizes, materials, greases, etc., and thus different noises. Therefore, a trained detection model for a certain type of bearing cannot be directly used for other new types of bearings.

[0003] Another problem with typical bearing noise detection models is that the detection completely depends on some key parameters of the bearing, such as fault frequencies (BPFI, BPFO, etc.) and relatively stable and accurate rotational speed information. However, in some cases, it is difficult to collect this parameter information, such as the lack of information about the bearing type or the lack of bearing speed information.

[0004] Current bearing detection algorithms need to collect a large amount of bearing data related to bearing information to retrain the model for new types of bearings. This not only wastes time and cost but also cannot support rapid expansion in more and more new applications. In addition, for the case where there is no bearing type and speed information, only professional personnel can perform manual diagnosis, or diagnosis can only be based on very simple diagnostic logics (such as RMS thresholds or trends), so the detection efficiency and detection accuracy are relatively low.

[0005] Therefore, there is a need to develop a general method and system that can perform fault judgment through automatic bearing noise detection without bearing information, so as to improve the efficiency of bearing noise detection by improving the generalization ability of the automatic bearing noise detection model. Summary of the Invention

[0006] One or more embodiments of the present invention provide a method for fault diagnosis through bearing noise detection. The method includes: collecting noise in bearing detection, where the noise includes bearing noise and operating condition noise; preprocessing the collected noise to obtain a first time-domain signal and a second frequency-domain signal; and inputting the first time-domain signal and the second frequency-domain signal into a bearing fault diagnosis model for fault diagnosis. Among them, the bearing fault diagnosis model includes a feature extraction model and a fault discrimination model. The feature extraction model respectively extracts features from the first time-domain signal and the second frequency-domain signal to obtain a first feature associated with time-domain impact peaks and a second feature associated with fault frequency peaks. And the fault discrimination model combines the first feature and the second feature and discriminates faults based on the combined features.

[0007] Among them, the preprocessing of the collected bearing vibration signal may include: performing band-pass filtering on the collected bearing vibration signal. The preprocessing of the collected noise may further include obtaining the time-domain waveform data of the filtered signal as the first time-domain signal, and performing Fourier transform on the filtered signal and obtaining the envelope spectrum data of the Fourier-transformed signal as the second frequency-domain signal. Optionally, the preprocessing may further include performing normalization processing on the first time-domain signal and the second frequency-domain signal respectively. Optionally, the preprocessing may further include resampling the normalized first time-domain signal and second frequency-domain signal.

[0008] Among them, the feature extraction model may include models for processing time-domain signals and frequency-domain signals respectively. For example, a first sub-model for processing the first time-domain signal and a second sub-model for processing the second frequency-domain signal.

[0009] The method may further include, based on the first time-domain signal, the first sub-model using a convolution kernel to extract first peak data and first average data respectively; and based on the second frequency-domain signal, the second sub-model using a convolution kernel to extract second peak data and second average data respectively. Among them, the first peak data represents the noise impact at a specific time in the time domain, and the first average data represents the average working condition noise in the time domain. Among them, the second peak data represents the noise impact at a specific spectrum in the frequency domain, and the second average data represents the average working condition noise in the frequency domain.

[0010] The method may further include: combining the first peak data and the first average data to obtain the first feature; and combining the second peak data and the second average data to obtain the second feature.

[0011] Among them, the feature extraction model may be a model based on a convolutional neural network (CNN), and the fault discrimination model may be a model based on a fully connected network.

[0012] Among them, the first time-domain signal may be a time-domain envelope, and the second frequency-domain signal may be a spectrum envelope.

[0013] The method may further include establishing a fault mode data set based on historical fault modes, and storing the discriminated fault modes to update the fault mode data set.

[0014] One or more embodiments of the present invention provide a system for fault diagnosis through bearing noise detection. The system includes a data collector, a processor, and a memory. The data collector is configured to collect the noise in bearing detection, and the noise includes bearing noise and operating condition noise. The processor is connected to the data collector and is configured to perform preprocessing on the collected noise to obtain a first time-domain signal and a second frequency-domain signal. The processor is further configured to input the first time-domain signal and the second frequency-domain signal into a bearing fault diagnosis model to discriminate bearing faults. Among them, the bearing fault diagnosis model includes a feature extraction model and a fault discrimination model. The processor is configured to enable the feature extraction model to respectively extract features from the first time-domain signal and the second frequency-domain signal to obtain a first feature associated with the time-domain impact peak and a second feature associated with the fault frequency peak; and merge the first feature and the second feature through the fault discrimination model, and discriminate the fault based on the merged features. The memory is configured to be connected to the processor and store the obtained fault identification result to update the original fault database.

[0015] One or more embodiments of the present invention provide a computer-readable storage medium including instructions that are executed by a computer to implement the above method for fault diagnosis through bearing noise detection.

[0016] Advantageously, the method and system for fault diagnosis through bearing noise detection disclosed by the present invention can achieve fault judgment through automatic bearing noise detection without bearing information, thereby improving the accuracy and efficiency of bearing noise detection by improving the generalization ability of the automatic bearing noise detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The system can be better understood with reference to the following description and in conjunction with the drawings. The components in the drawings are not to scale, but the emphasis is on illustrating the principles of the present invention. In addition, in the drawings, like or identical reference numerals represent like or identical elements.

[0018] Figure 1 Schematically shows a flowchart of a method for fault identification by detecting bearing noise according to one or more embodiments of the present invention.

[0019] Figure 2 Schematically shows a flowchart of a method for preprocessing the collected bearing noise according to one or more embodiments of the present invention.

[0020] Figure 3 Schematically shows a simplified diagram of a bearing fault diagnosis model according to one or more embodiments of the present invention.

[0021] Figure 4An example of a network model of a bearing fault diagnosis model according to the present invention is schematically shown. Detailed implementation manners

[0022] It should be understood that the following description of the embodiments is only for illustrative purposes and not restrictive. The division of examples in the functional blocks, modules or units shown in the drawings should not be construed as indicating that these functional blocks, modules or units must be implemented as physically separate units. The functional blocks, modules or units shown or described can be implemented as separate units, circuits, chips, functions, modules or circuit elements. One or more functional blocks or units can also be implemented in a common circuit, chip, circuit element or unit.

[0023] Figure 1 A flowchart of a method for fault identification by detecting bearing noise according to one or more embodiments of the present invention is schematically shown.

[0024] Reference Figure 1 , at S101, the noise in the bearing detection can be collected through sensors, such as one or more vibration sensors. The noise includes bearing noise caused by bearing faults and also includes operating condition noise. At S102, preprocessing can be performed on the noise signal collected through the sensors. At S103, the preprocessed signal is input into the bearing fault diagnosis model for diagnosis. Optionally, at S104, the fault diagnosis result is stored and / or output. For example, the output fault diagnosis result can be displayed to the operator / user through a display device or an alarm sound can be emitted through an alarm device to alert the operator / user that a bearing fault has been detected, or the fault diagnosis result can be stored and displayed to the operator / user when the operator / user calls the result.

[0025] Figure 2 A flowchart of an exemplary method for preprocessing the collected bearing noise data according to one or more embodiments of the present invention is schematically shown.

[0026] First, at S201, the collected noise signal can be filtered. For example, the noise signal can be filtered by using a band-pass filter with a bandwidth of 500 Hz - 10,000 Hz. In actual operation, the bandwidth range of the band-pass filter can be adjusted according to the actual situation. For example, the bandwidth of the band-pass filter can be adjusted based on at least one of the operating condition speed and load.

[0027] Next, the filtered noise signal is processed in two parallel paths. One path of processing is to process the filtered signal in the time domain, and the other path of processing is to process the filtered signal in the frequency domain. For example, at S202, the time-domain waveform data (i.e., time series data) of the filtered signal can be obtained as the first time-domain signal. For example, at S203, the Fourier transform can be performed on the filtered signal to convert the filtered time-domain signal into a frequency-domain signal, and the spectral envelope data (such as ENV data) of the frequency-domain signal is calculated as the second frequency-domain signal.

[0028] At S204, the time-domain waveform data (time-domain envelope data) and the frequency-domain spectral envelope data can be normalized. And, according to the length requirement of the input signal for the bearing fault diagnosis model, the time-domain waveform data and the frequency-domain spectral envelope data are resampled at a constant length respectively. For example, the usually preferred constant length is 128*128. However, those skilled in the art can understand that other constant lengths can be adaptively adopted according to the actual needs and the specific design of the bearing fault diagnosis model. Subsequently, the resampled time-domain waveform data and the frequency-domain spectral envelope data are input into the bearing fault diagnosis model for fault identification. Those skilled in the art can understand that the above steps of preprocessing the data can be executed sequentially and / or in parallel.

[0029] The following will refer to Figure 3 and Figure 4 to introduce the bearing fault diagnosis model designed in some embodiments of the present invention. The bearing fault diagnosis model in some embodiments of the present invention is designed based on a deep neural network. Figure 3 is a schematic diagram of the bearing fault diagnosis model according to one or more embodiments of the present invention. As Figure 3As shown, the bearing fault diagnosis model of the present invention can be divided into two parts. The first part is a feature extraction model based on a convolutional neural network (CNN), and the second part is a fault discrimination model based on a fully connected network. In addition, the bearing fault diagnosis model of this embodiment simultaneously considers two types of data inputs, namely time-domain envelope signal data and frequency-domain envelope spectrum data, and finally identifies faults by automatically extracting features in two modes: time-domain impact peaks and frequency-domain fault frequency peaks. Specifically, using the pooling technique in the convolutional network and selecting a specific pooling function (such as the maximum function), the local peaks of the spectrum after convolution are extracted, and stable peak features are extracted through multi-level convolution. The peak itself represents an impact at a specific time or specific frequency spectrum, and this impact is often caused by partial faults of the equipment (for example, bearing faults). Generally, normal bearings do not have obvious impacts. In addition, other pooling functions (such as the average value) are simultaneously selected to extract the mean value of the spectrum after convolution, which represents the average energy in the time domain and frequency domain. This often represents the average working condition vibration and / or noise, and as the bearing fault becomes more obvious, the overall noise during the detection process will also become larger. Through the feature extraction model, features related to fault impacts (such as local peaks) and features related to the average noise / vibration of the bearing (such as local means) extracted from two different envelope lines in the time domain / frequency domain are used as effective parameters for judging the bearing health state. Usually, under good bearing conditions, the average vibration of the bearing is very small, and the features related to fault impacts are very obvious in both the frequency domain and the time domain, and are easy to detect; however, detection based solely on this feature is also prone to premature discrimination, overestimating the bearing fault and resulting in unnecessary waste, such as excessive defect rates or additional maintenance and detection costs. However, as the health state of the bearing deteriorates, not only do local impacts increase, but the average noise of the bearing also increases. These fault features will show very different performances in the time domain and frequency domain. Sometimes they are not obvious in the time domain but will stand out in the frequency domain. In addition, due to complex factors, the results of detection based on a fixed threshold using the above scheme are not ideal. In the above embodiments of the present invention, based on algorithm models such as neural networks, it is possible to learn from a large amount of data of labeled data in different states, and finally achieve model convergence and a good recognition rate. It should be noted that although the above embodiments of the present invention mainly use neural networks as examples, these examples are not used for limitation. Any algorithm model similar to a neural network and implementing the solution of the present invention based on the spirit and ideas shown in the above embodiments of the present invention should be included within the scope of the present invention.

[0030] Reference Figure 3 , the above reference Figure 2The preprocessed data described is input as input data into the first part of the bearing fault diagnosis model, namely the feature extraction model based on the CNN network. The feature extraction model of the present invention includes two independent branch sub-models to separately process the input time-domain envelope and frequency-domain envelope. The two independent branch sub-models can respectively achieve the feature extraction of time-domain signals and frequency-domain signals.

[0031] For example, Figure 3 (The left half) schematically shows the first sub-model for feature extraction of time-domain signals, which uses a convolutional kernel to parallelly extract local peaks and averages from the time-domain envelope spectrum line. Among them, the local peaks represent possible fault impacts, and the local averages represent the noise level under the working conditions. The peaks and averages after convolution are the self-extracted key features for fault detection. For the purpose of exemplarily explaining the principle, Figure 3 the multi-layer network structure is omitted, and only the network structure is schematically shown in the form of a simple diagram. In actual operation, a multi-layer network can often be designed according to the actual situation to ensure the stability of extracting various peaks and reduce the sensitivity of the peaks to the spectral position. The local peaks and averages obtained through the multi-layer network are combined to calculate the first feature associated with the time-domain impact peak of the first sub-model.

[0032] Similarly, Figure 3 (The right half) schematically shows the second sub-model for feature extraction of frequency-domain signals, which also uses a convolutional kernel to extract local peaks and averages from the frequency-domain envelope spectrum line. And, the local peaks and averages obtained through the multi-layer network are combined to calculate the second feature associated with the fault frequency peak of the second sub-model.

[0033] Next, the first feature of the first sub-model and the second feature of the second sub-model calculated are input into the fault discrimination model based on the fully connected network and feature combination is performed. And, by comprehensively considering the fault impact feature and the working condition noise feature, finally, a multi-layer fully connected network is used to achieve fault discrimination based on features. The multi-layer fully connected network implements a special classifier, which can correctly distinguish normal bearings and faulty bearings according to faults and working condition noise. Faulty bearings often have more impact features and relatively large working condition noise.

[0034] Figure 4A deep neural network model showing an exemplary bearing fault diagnosis model of the present invention is presented. This model consists of two parts. The first part is a feature extraction model based on a convolutional neural network (CNN), and the second part is a fault discrimination model based on a fully connected network. Among them, the feature extraction model based on CNN includes two independent branch sub-models to separately process the input time-domain envelope and frequency-domain envelope. The two independent branch sub-models can respectively and parallelly implement feature extraction of the time-domain signal and frequency-domain signal of the bearing vibration signal.

[0035] Among them, the first sub-model for feature extraction of the time-domain signal may include multiple convolutional layers, such as the convolutional layers conv1D_4:Conv1D, conv1D_5:Conv1D, conv1D_6:Conv1D for extracting local peaks through convolution, and the convolutional layers conv1D_7:Conv1D, conv1D_8:Conv1D, conv1D_9:Conv1D for extracting local averages through convolution. The convolutional layer uses a convolutional kernel to perform convolutional calculations on a local area of the input signal or feature to extract the required key feature information. The first sub-model also includes multiple pooling layers. The main role of the pooling layer is to reduce the number of parameters and the number of original features. Although the convolutional layer has largely reduced the number of connections in the neural network, for the neurons in the feature map group, the number has not decreased significantly, and the subsequent input dimension is still relatively high, which is prone to overfitting. Therefore, the convolutional layer and the pooling layer are usually used in combination to effectively reduce the feature dimension. In Figure 4 the exemplary model shown, the first sub-model includes max-pooling layers for extracting local peaks, such as max_pooling1d_3:MaxPooling1D, max_pooling1d_4:maxPooling1D, max_pooling1d_5:MaxPooling1D, and average-pooling layers for extracting averages, such as average_pooling1d_2:AveragePooling1D, average_pooling1d_3:AveragePooling1D, average_pooling1d_4:AveragePooling1D. Those skilled in the art can understand that Figure 4 only an exemplary multi-layer convolutional layer and pooling layer structure is shown, and in actual operation, more or fewer convolutional layers and pooling layers can be designed according to different requirements.

[0036] Next, for example, the data representing local peaks output from the max pooling layer max_pooling1d_5:MaxPooling1D is flattened into one-dimensional peak data by the flattening layer flattern_3:Flattern, and the data representing the average value output from the average pooling layer average_pooling1d_4:AveragePooling1D is flattened into one-dimensional average value data by the flattening layer flattern_4:Flattern. The flattened one-dimensional peak data and one-dimensional average value data are combined in the concatenation layer concatenate_2:Concatenate to extract the first feature of the first sub-model. This first feature is input into the concatenation layer concatenate_2:Concatenate after passing through the dense layer dense_2:Dense to be combined with the second feature extracted from the second sub-model for feature extraction of the frequency domain signal.

[0037] Similarly, the second sub-model for feature extraction of the frequency domain signal can also include multiple convolutional layers, such as the convolutional layers conv1D_1:Conv1D and conv1D_2:Conv1D for extracting local peaks through convolution, and the convolutional layer conv1D_3:Conv1D for extracting local average values through convolution. The second sub-model can also include multiple pooling layers, such as the max pooling layers max_pooling1d_1:MaxPooling1D and max_pooling1d_2:maxPooling1D for extracting local peaks, and the average pooling layer average_pooling1d_1:AveragePooling1D for extracting average values.

[0038] For example, the frequency-domain signal entering the second sub-model goes through the convolutional layer conv1d_1:Conv1D and the pooling layer max_pooling1d_1:MaxPooling1D in one branch, and then goes through conv1d_2:Conv1D and max_pooling1d_2:MaxPooling1D. The data representing local peaks output from the max pooling layer max_pooling1d_2:MaxPooling1D is flattened into one-dimensional peak data by the layer flattern_1:Flattern. At the same time, the frequency-domain signal entering the second sub-model goes through the convolutional layer conv1D_3:Conv1D and the average pooling layer average_pooling1d_1:AveragePooling1D in another branch. The data representing the average value output from the average pooling layer does not go through additional convolutional and pooling layers, but is directly flattened into one-dimensional average value data by the layer flattern_2:Flattern. Subsequently, the flattened one-dimensional peak data and one-dimensional average value data are merged in the merging layer concatenate_1:Concatenate to extract the second feature of the second sub-model. This second feature is input into the merging layer concatenate_3:Concatenate after passing through the dense layer dense_1:Dense to be merged with the first feature extracted from the first sub-model.

[0039] It can be seen that the number of convolutional and pooling layers in the two sub-models can be different, and the number of max pooling and average pooling layers in the same sub-model can also be different. Those skilled in the art can understand that the number of these convolutional and pooling layers varies according to specific actual operations. Similarly, the number of merging, flattening, and dense layers can also vary according to specific requirements, or some of these layers can be omitted.

[0040] The first feature extracted by the first sub-model and the second feature extracted by the second sub-model are input into the fault discrimination model based on a fully connected network and are first merged in the merging layer concatenate_3:Concatenate. By comprehensively considering the fault impact feature and the working condition noise feature, a feature-based fault discrimination is further realized through a multi-layer fully connected network such as including dense_3:Dense, dropout_1:Dropout, dense_4:Dense, and dense_5:Dense.

[0041] The bearing fault diagnosis model designed by the above embodiment of the present invention can realize an automatic bearing fault diagnosis model without any bearing information and speed information. When constructing the bearing fault diagnosis model of the present invention, typical fault modes can be collected from all historical fault databases to form a fault mode original data set for training the fault diagnosis model of the above embodiment of the present invention. The fault diagnosis model of the above embodiment of the present invention diagnoses bearing faults only by comparing the vibration (noise) waveform shape or specific pattern in the bearing detection with the typical fault mode, without the need for any detailed bearing information and speed information. In most cases, the waveform shape and specific pattern are much more important than the absolute amplitude value. During training, the original data set needs to be preprocessed before the fault mode original data set is input into the neural network model of the present invention. This preprocessing method is combined with the above Figure 2 The preprocessing method described above is the same and will not be repeated here. In addition, in order to expand the training data set, the spectrum envelope curve data and time series curve data of the data can be randomly extracted from the historical fault database, and white noise can be randomly added to generate more training data sets. At the same time, the fault type data detected in the actual detection can also be continuously stored in the original data set to expand and update the original training data set.

[0042] The special bearing fault diagnosis model structure designed in the present invention can provide better generalization ability. The model structure can be directly used for different bearing types, grease types, seal types, etc. In addition to being used in bearing manufacturing, the fault diagnosis model structure of the present invention can also be used in the fault diagnosis process of any mechanical structure including bearings to achieve automatic diagnosis. The method of the present invention is also suitable for building a model in a public cloud and connecting to any client to diagnose bearing faults without bearing types. For example, the bearing noise detection method of the present invention can be used for the main bearing fault diagnosis in wind turbines without the need for turbine type and bearing information, thereby greatly saving computing time and cost, and thus improving detection efficiency. The method of the present invention can also be extended to different applications, such as gearbox fault detection, imbalance fault detection, etc.

[0043] One or more embodiments of the present invention also provide a system for bearing noise detection. The system includes a data collector, such as a vibration sensor. The data collector is configured to collect vibration (noise) signals during bearing detection. The system of the present invention may also include a processor connected to the data collector. The processor of the present invention as a whole may be a microprocessor, an application specific integrated circuit (ASIC), a system on a chip (SoC), a computing device, a portable mobile computing device (such as a tablet computer or a mobile phone), etc. The processor may be configured to execute the following methods: preprocess the collected vibration (noise) signals to obtain a time domain signal and a frequency domain signal; and input the time domain signal and the frequency domain signal into a bearing fault diagnosis model. Among them, the bearing fault diagnosis model includes a feature extraction model and a fault discrimination model. The feature extraction model respectively extracts features from the time domain signal and the frequency domain signal to obtain a first feature associated with the time domain impact peak and a second feature associated with the fault frequency peak. The fault discrimination model combines the first feature and the second feature to obtain a fault identification result. The system may also include a memory connected to the processor. The memory may store an original fault data set and may store the obtained fault identification result to update the original fault data set.

[0044] Any one or more of the processors, memories, or systems described herein include computer-executable instructions that may be compiled or interpreted from computer programs created using various programming languages and / or technologies. Generally, a processor (such as a microprocessor) receives instructions from, for example, a memory, a computer-readable medium, etc. and executes the instructions. The processor includes a non-transitory computer-readable storage medium capable of executing the instructions of a software program. The computer-readable medium may be, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof.

[0045] The description of the embodiments has been presented for purposes of illustration and description. Appropriate modifications and variations of the embodiments may be made in light of the above description and may be obtained by practicing the methods. For example, unless otherwise indicated, one or more of the described methods may be performed by a suitable combination of devices and / or systems. The methods may be performed by: using one or more logic devices (such as a processor) in combination with one or more additional hardware elements (such as a storage device, a memory, a circuit, a hardware network interface, etc.) to execute stored instructions. The methods and associated actions may also be performed in parallel and / or simultaneously in various orders other than the order described in this application. The system is exemplary in nature and may include additional elements and / or omit elements. The subject matter of the present disclosure includes all novel and non-obvious combinations of the various methods and system configurations and other features, functions, and / or properties disclosed.

[0046] As used in this application, an element or step recited in the singular and preceded with the word "a" or "an" should be understood as not excluding a plurality of said elements or steps, unless such exclusion is stated. Additionally, a reference to "one embodiment" or "an example" of the present disclosure is not intended to be construed as excluding the existence of additional embodiments that also incorporate the recited features. The present invention has been described above with reference to specific embodiments. However, those of ordinary skill in the art will understand that various modifications and changes can be made thereto without departing from the broader spirit and scope of the invention as set forth in the appended claims.

Claims

1. A method for fault diagnosis through bearing noise detection, comprising: Collecting the noise in bearing detection, where the noise includes bearing noise and operating condition noise; Preprocessing the collected noise to obtain a first time-domain signal and a second frequency-domain signal; And Inputting the first time-domain signal and the second frequency-domain signal into a bearing fault diagnosis model; Wherein, the bearing fault diagnosis model includes a feature extraction model and a fault discrimination model, and the feature extraction model respectively extracts features from the first time-domain signal and the second frequency-domain signal to obtain a first feature associated with time-domain impact peaks and a second feature associated with fault frequency peaks; And the fault discrimination model combines the first feature and the second feature and discriminates faults based on the combined features.

2. The method according to claim 1, wherein The feature extraction model includes a first sub-model for processing the first time-domain signal to obtain the first feature and a second sub-model for processing the second frequency-domain signal to obtain the second feature.

3. The method according to claim 2, wherein Based on the first time-domain signal, the first sub-model respectively extracts first peak data and first average data, where the first peak data represents the noise impact in the time domain, and the first average data represents the average operating condition noise in the time domain; and Based on the second frequency-domain signal, the second sub-model respectively extracts second peak data and second average data, where the second peak data represents the noise impact in the frequency domain, and the second average data represents the average operating condition noise in the frequency domain.

4. The method according to claim 3, further comprising: Combining the first peak data and the first average data to obtain the first feature; And Combining the second peak data and the second average data to obtain the second feature.

5. The method according to any one of claims 1-4, wherein, The feature extraction model is a model based on a convolutional neural network (CNN), and the fault discrimination model is a model based on a fully connected network.

6. The method according to any one of claims 1-4, wherein The first time-domain signal is the time-domain envelope of the signal, and the second frequency-domain signal is the spectral envelope of the signal.

7. The method according to any one of claims 1-4, further comprising establishing a fault mode data set based on historical fault modes and storing the discriminated fault modes to update the fault mode data set.

8. The method according to any one of claims 1-4, wherein the preprocessing includes: Performing band-pass filtering on the collected noise; Obtaining the time-domain waveform data of the filtered signal as the first time-domain signal; Performing Fourier transform on the filtered signal and obtaining the envelope spectrum data of the Fourier-transformed signal as the second frequency-domain signal; Performing normalization processing on the first time-domain signal and the second frequency-domain signal respectively; And Performing resampling on the normalized first time-domain signal and the second frequency-domain signal.

9. A system for fault diagnosis through bearing noise detection, comprising: A data collector configured to collect the noise in bearing detection, where the noise includes bearing noise and operating condition noise; A processor connected to the data collector, and the processor is configured to: Preprocess the collected noise to obtain a first time-domain signal and a second frequency-domain signal; and Input the first time-domain signal and the second frequency-domain signal into a bearing fault diagnosis model; wherein the bearing fault diagnosis model includes a feature extraction model and a fault discrimination model, the feature extraction model respectively extracts features from the first time-domain signal and the second frequency-domain signal to obtain a first feature associated with a time-domain impact peak and a second feature associated with a fault frequency peak; and the fault discrimination model combines the first feature and the second feature and discriminates faults based on the combined features; and a memory configured to be connected to the processor and store the discriminated fault mode to update the fault mode dataset.

10. A computer-readable storage medium including instructions that are executed by a computer to implement the method according to any one of claims 1-8.

Images