Low-speed heavy-duty bearing fault feature mining method and system based on broadband mode decomposition

Abstract

The invention relates to the technical field of bearing fault diagnosis, in particular to a low-speed heavy-duty bearing fault feature mining method and system based on broadband modal decomposition, and the method comprises the steps: collecting an original signal through a multi-channel vibration sensor, and carrying out the preprocessing and broadband modal decomposition to extract a target energy distribution matrix; respectively calculating a first matching degree, a second matching degree and a third matching degree by utilizing the typical energy distribution matrix, the reference energy fluctuation interval set and a deep learning model; and combining the multi-dimensional weight set with a comprehensive matching formula to generate a target matching confidence coefficient set, and finally identifying a target fault mode corresponding to the maximum matching confidence coefficient. The precision, the robustness and the real-time performance of fault diagnosis of the low-speed heavy-duty bearing under the complex working condition can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bearing fault diagnosis, and specifically relates to a low-speed heavy-load bearing fault feature mining method and system based on wide-frequency modal decomposition. BACKGROUND

[0002] The existing low-speed heavy-load bearing fault diagnosis method still has deficiencies in the accuracy of signal decomposition, the comprehensiveness of feature extraction, and the robustness and generalization ability of the diagnosis model, and it is difficult to meet the high-precision diagnosis demand under complex working conditions.

[0003] In the prior art, a bearing fault diagnosis method based on single-component variational modal decomposition (VMD) is proposed in the patent with the publication number CN117740381B. The method extracts high-order features through a stacked autoencoder (SAE) and constructs a classification model combined with a random forest (RF). This method can effectively avoid modal aliasing and end effects, but the single-component VMD only decomposes a single mode, which may not fully extract multi-scale feature information in the wide-frequency signal, especially under low signal-to-noise ratio conditions. In addition, although the combined model of SAE and RF has a certain robustness, it may have insufficient generalization ability when dealing with complex nonlinear features, and the discrimination degree for multiple categories of faults needs to be further improved.

[0004] Another prior art, the patent with the publication number CN114813131B, proposes a low-speed heavy-load bearing fault recognition method based on filter decomposition and BP neural network. The method extracts feature quantities of the first three component signals by filtering and decomposing the signal, and realizes fault recognition through BP neural network after dimensionality reduction using distance evaluation technology. However, this technical solution mainly focuses on the first three component signals, which may cause some important fault information contained in high-frequency or low-frequency components to be ignored, thereby affecting the comprehensiveness of feature extraction. At the same time, the BP neural network is sensitive to the initial weight, and is prone to fall into local optimum, and the training efficiency on large-scale data sets is low, which may lead to insufficient stability and real-time performance of the diagnosis results.

[0005] The above problems show that the existing low-speed heavy-load bearing fault diagnosis method still has obvious deficiencies in the multi-scale feature mining ability of signal decomposition, the comprehensiveness of feature extraction, and the robustness and real-time performance of the diagnosis model. Therefore, there is an urgent need for a technical solution that can fully extract multi-scale feature information from the signal, optimize the feature extraction process, and improve the performance of the diagnosis model, to meet the high-precision, high-robustness, and high-real-time performance requirements of low-speed heavy-load bearing fault diagnosis under complex working conditions. The present application aims to solve the problems in the prior art by introducing wide-frequency modal decomposition technology and combining optimized feature extraction and classification model, to provide more reliable technical support for low-speed heavy-load bearing fault diagnosis. SUMMARY

[0006] The application provides a low-speed heavy-load bearing fault feature mining method and system based on wide-frequency modal decomposition, which aims to improve the precision, robustness and real-time performance of low-speed heavy-load bearing fault diagnosis under complex working conditions. To achieve the above purpose, the low-speed heavy-load bearing fault feature mining method based on wide-frequency modal decomposition provided by the application comprises the following steps: collecting the original vibration signal of the bearing through a pre-constructed multi-channel vibration sensor; preprocessing the original vibration signal to obtain an initial vibration signal; performing wide-frequency modal decomposition on the initial vibration signal to obtain a plurality of groups of wide-frequency modal components; extracting the energy distribution of the plurality of groups of wide-frequency modal components according to a pre-set frequency interval to obtain a target energy distribution matrix; sequentially extracting bearing fault modes in a pre-set bearing fault mode set to obtain a typical energy distribution matrix of the bearing fault mode; calculating the first matching degree of the target energy distribution matrix and the typical energy distribution matrix according to a pre-set energy matching rule to obtain a first matching degree set; receiving a plurality of sets of reference energy fluctuation intervals input by a user according to the bearing fault mode, identifying a plurality of sets of target energy fluctuation intervals of the target energy distribution matrix, calculating the second matching degree of the plurality of sets of target energy fluctuation intervals and the plurality of sets of reference energy fluctuation intervals to obtain a second matching degree set; inputting the target energy distribution matrix into a pre-constructed target deep learning model to obtain a third matching probability value set; calculating the target matching confidence set by using a pre-constructed comprehensive matching formula according to a pre-set multi-dimensional weight set, the first matching degree set, the second matching degree set and the third matching probability value set; identifying the maximum target matching confidence in the target matching confidence set and identifying the target bearing fault mode corresponding to the maximum target matching confidence.

[0007] Optionally, the energy distribution extraction of the plurality of groups of wide-frequency modal components according to the pre-set frequency interval to obtain the target energy distribution matrix comprises: performing one-time segmentation on the plurality of groups of wide-frequency modal components according to the frequency interval to obtain low-frequency modal components, medium-frequency modal components and high-frequency modal components; performing two-time segmentation on the low-frequency modal components, the medium-frequency modal components and the high-frequency modal components according to a pre-set time window to obtain a plurality of groups of time window low-frequency modal sequences, a plurality of groups of time window medium-frequency modal sequences and a plurality of groups of time window high-frequency modal sequences; sequentially performing amplitude peak value extraction in the plurality of groups of time window low-frequency modal sequences, the plurality of groups of time window medium-frequency modal sequences and the plurality of groups of time window high-frequency modal sequences to obtain a plurality of groups of low-frequency energy point sequences, a plurality of groups of medium-frequency energy point sequences and a plurality of groups of high-frequency energy point sequences; performing energy fitting on the plurality of groups of low-frequency energy point sequences, the plurality of groups of medium-frequency energy point sequences and the plurality of groups of high-frequency energy point sequences to obtain a target low-frequency energy distribution, a target medium-frequency energy distribution and a target high-frequency energy distribution; and performing matrix splicing on the target low-frequency energy distribution, the target medium-frequency energy distribution and the target high-frequency energy distribution to obtain the target energy distribution matrix.

[0008] Optionally, the acquiring the typical energy distribution matrix of the bearing fault mode comprises: acquiring an energy distribution matrix set of the bearing fault mode; extracting energy distribution matrices in the energy distribution matrix set in sequence, receiving user's cutting of the energy distribution matrices according to the signal monitoring time length and the signal monitoring frequency band, and obtaining a set of to-be-mean energy distribution matrices; determining a signal monitoring time point and a signal monitoring frequency point according to the signal monitoring time length and the signal monitoring frequency band, constructing an energy calibration point array according to the signal monitoring time point and the signal monitoring frequency point, extracting energy calibration points in the energy calibration point array in sequence, identifying a corresponding set of to-be-mean energy values in the set of to-be-mean energy distribution matrices according to the energy calibration points, calculating an average energy value of the set of to-be-mean energy values, determining a three-dimensional energy point according to the average energy value and the energy calibration point, and collecting three-dimensional energy points of each energy calibration point to obtain a set of three-dimensional energy points; and fitting the set of three-dimensional energy points to obtain the typical energy distribution matrix.

[0009] Optionally, the energy matching rule is as follows: wherein, represents a first matching degree of the typical energy distribution matrix of the i-th bearing fault mode and the target energy distribution matrix, represents a number of energy calibration points in the energy calibration point array, represents an energy value of the p-th energy calibration point in the target energy distribution matrix, represents an energy value of the p-th energy calibration point in the i-th typical energy distribution matrix of the bearing fault mode, and represents an absolute value symbol.

[0010] Optionally, the identifying the set of target energy fluctuation intervals of the target energy distribution matrix comprises: extracting signal monitoring frequencies in the signal monitoring frequency points in sequence, and identifying an energy change curve of the signal monitoring frequencies in the target energy distribution matrix; identifying a set of target waveform periods, a set of target waveform amplitudes, and a set of target waveform steepnesses of the energy change curve; calculating a target average waveform period, a target average waveform amplitude, and a target average waveform steepness according to the set of target waveform periods, the set of target waveform amplitudes, and the set of target waveform steepnesses, respectively, to obtain a set of target energy fluctuation intervals; and collecting the sets of target energy fluctuation intervals of each signal monitoring frequency to obtain a plurality of sets of target energy fluctuation intervals.

[0011] Optionally, the calculating the second matching degree of the plurality of sets of target energy fluctuation intervals and the plurality of sets of reference energy fluctuation intervals to obtain a second matching degree set comprises: calculating the second matching degree according to the plurality of sets of target energy fluctuation intervals and the plurality of sets of reference energy fluctuation intervals, wherein represents the second matching degree of the plurality of sets of target energy fluctuation intervals and the plurality of sets of reference energy fluctuation intervals of the i-th bearing fault mode, represents a period coefficient, represents an amplitude coefficient, represents a steepness coefficient, represents the number of signal monitoring frequencies in the signal monitoring frequency point, represents a reference waveform period in the q-th set of reference energy fluctuation intervals, represents a target average waveform period in the q-th set of target energy fluctuation intervals, represents a reference waveform amplitude in the q-th set of reference energy fluctuation intervals, represents a target average waveform amplitude in the q-th set of target energy fluctuation intervals, represents a reference waveform steepness in the q-th set of reference energy fluctuation intervals, and represents a target average waveform steepness in the q-th set of target energy fluctuation intervals; and the second matching degrees corresponding to each bearing fault mode are collected to obtain the second matching degree set.

[0012] Optionally, the inputting the target energy distribution matrix into a pre-constructed target deep learning model to obtain a third matching probability value set comprises: extracting an amplitude sequence in an energy change curve of the signal monitoring frequency according to a preset amplitude sampling interval to obtain an amplitude sequence set corresponding to different signal monitoring frequencies; and inputting the amplitude sequence set into the target deep learning model to obtain the third matching probability value set, wherein an input layer of the target deep learning model comprises a first amplitude sequence input node, a second amplitude sequence input node, and an n-th amplitude sequence input node, and an output layer of the target deep learning model comprises a first fault mode probability value output node, a second fault mode probability value output node, and an n-th fault mode probability value output node.

[0013] Optionally, the obtaining manner of the multi-dimensional weight set comprises: sequentially extracting a training energy distribution matrix set from a plurality of preset training energy distribution matrix sets; sequentially extracting a training energy distribution matrix from the training energy distribution matrix set; calculating a first training matching degree set of the training energy distribution matrix according to the typical energy distribution matrix and the energy matching rule, identifying a maximum first training matching degree in the first training matching degree set, identifying a first predicted failure mode corresponding to the maximum first training matching degree, and obtaining a first predicted failure mode set; identifying a plurality of training energy fluctuation interval sets of the training energy distribution matrix, calculating a second training matching degree set according to the plurality of reference energy fluctuation interval sets and the plurality of training energy fluctuation interval sets, identifying a maximum second training matching degree in the second training matching degree set, identifying a second predicted failure mode corresponding to the maximum second training matching degree, and obtaining a second predicted failure mode set; predicting a third training probability value set of the training energy distribution matrix by using the target deep learning model, identifying a maximum third training probability value in the third training probability value set, identifying a third predicted failure mode corresponding to the maximum third training probability value, and obtaining a third predicted failure mode set; obtaining a real failure mode of the training energy distribution matrix to obtain a real failure mode set; calculating sequence consistency of the first predicted failure mode set, the second predicted failure mode set and the third predicted failure mode set with the real failure mode set respectively to obtain first sequence consistency, second sequence consistency and third sequence consistency; determining whether the first sequence consistency, the second sequence consistency and the third sequence consistency are stable; if the first sequence consistency, the second sequence consistency and the third sequence consistency are not stable, returning to the step of sequentially extracting a training energy distribution matrix set from a plurality of preset training energy distribution matrix sets; if the first sequence consistency, the second sequence consistency and the third sequence consistency are stable, normalizing the first sequence consistency, the second sequence consistency and the third sequence consistency to obtain a multi-dimensional weight set.

[0014] Optionally, the judging whether the first sequence consistency, the second sequence consistency and the third sequence consistency tend to be stable comprises: obtaining a first sequence consistent sequence, a second sequence consistent sequence and a third sequence consistent sequence; drawing a first sequence consistent curve, a second sequence consistent curve and a third sequence consistent curve according to the first sequence consistent sequence, the second sequence consistent sequence and the third sequence consistent sequence respectively; identifying a first end slope, a second end slope and a third end slope of the first sequence consistent curve, the second sequence consistent curve and the third sequence consistent curve; judging whether the first end slope, the second end slope and the third end slope are all less than a preset end slope threshold; if the first end slope, the second end slope and the third end slope are not all less than the end slope threshold, the first sequence consistency, the second sequence consistency and the third sequence consistency do not tend to be stable; if the first end slope, the second end slope and the third end slope are all less than the end slope threshold, the first sequence consistency, the second sequence consistency and the third sequence consistency tend to be stable. ; in, The target energy distribution matrix represents the target matching confidence level for the i-th bearing failure mode. , and These represent the first fusion weight, the second fusion weight, and the third fusion weight, respectively. This represents the degree of matching between the target energy distribution matrix in the first matching set and the typical energy distribution matrix of the i-th bearing failure mode. This represents the matching degree between the target energy fluctuation interval set in the second matching degree set and the energy fluctuation interval set of the i-th bearing failure mode. The probability value of the failure mode of the target energy distribution matrix in the third matching probability value set being the i-th bearing failure mode; the target bearing failure mode identification module is used to identify the maximum target matching confidence in the target matching confidence set and identify the target bearing failure mode corresponding to the maximum target matching confidence.

[0015] To address the aforementioned problems, the present invention also provides an electronic device comprising: a memory storing at least one instruction; and a processor executing the instruction stored in the memory to implement the aforementioned method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition.

[0016] To address the aforementioned problems, the present invention also provides a computer-readable storage medium storing at least one instruction, which is executed by a processor in an electronic device to implement the aforementioned method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition.

[0017] To address the problems described in the background art, this invention requires first obtaining the target energy distribution matrix, then using different judgment methods to determine the bearing failure mode of the target energy distribution matrix, then fusing the various judgment results to calculate the target matching confidence set, and finally identifying the maximum target matching confidence score in the target matching confidence set and the target bearing failure mode corresponding to the maximum target matching confidence score. When obtaining the target energy distribution matrix, it is necessary to first collect the original vibration signal of the bearing using a pre-constructed multi-channel vibration sensor, then preprocess the original vibration signal to obtain the initial vibration signal. At this point, broadband mode decomposition is required on the initial vibration signal to obtain multiple sets of broadband mode components. Since different proportions of energy distribution extraction are required, it is necessary to first segment the multiple sets of broadband mode components according to a preset frequency range, and then perform energy distribution extraction to obtain the target energy distribution matrix. At this point, the bearing failure mode can be determined. First, the bearing failure modes need to be extracted sequentially from the preset bearing failure mode set. In the first judgment method... The invention employs a three-step process: First, it obtains a typical energy distribution matrix of the bearing fault mode. Then, it calculates the first matching degree between the target energy distribution matrix and the typical energy distribution matrix according to a preset energy matching rule, resulting in a first matching degree set. In the second method, it receives multiple sets of reference energy fluctuation intervals input by the user based on the bearing fault mode, identifies multiple sets of target energy fluctuation intervals of the target energy distribution matrix, and finally calculates the second matching degree between these sets of target energy fluctuation intervals and the multiple sets of reference energy fluctuation intervals, resulting in a second matching degree set. In the third method, it inputs the target energy distribution matrix into a pre-constructed target deep learning model to obtain a third matching probability value set. After completing these three calculations, it calculates the target matching confidence set using a pre-constructed comprehensive matching formula based on a preset multi-dimensional weight set, the first matching degree set, the second matching degree set, and the third matching probability value set. Finally, it identifies the maximum target matching confidence score within the target matching confidence score set and identifies the target bearing fault mode corresponding to the maximum target matching confidence score. Therefore, this invention improves the accuracy, robustness, and real-time performance of bearing fault diagnosis. Attached Figure Description

[0018] Fig. 1 This is a flowchart illustrating a method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition, provided in an embodiment of the present invention. Fig. 2 This is a schematic diagram of the target energy distribution matrix generation process provided in an embodiment of the present invention; Fig. 3 This is a functional block diagram of a low-speed heavy-load bearing fault feature mining system based on broadband mode decomposition provided in an embodiment of the present invention. Detailed Implementation

[0019] This invention provides a method and system for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition. Its core lies in achieving accurate identification of fault modes in low-speed heavy-load bearings through multi-channel vibration signal acquisition, signal preprocessing, broadband mode decomposition, energy distribution extraction, and multi-level matching degree calculation. (See attached figures.) Figs. 1 to 3 The specific implementation process of the present invention will be described in detail.

[0020] like Fig. 1 As shown, the implementation process of the method of the present invention includes several key steps. First, the raw vibration signal of the bearing is acquired using a pre-constructed multi-channel vibration sensor. These sensors are arranged at key locations on the bearing to ensure comprehensive capture of vibration information generated during bearing operation. To improve signal quality, vibration signal enhancement technology, noise suppression algorithms, and blind source separation methods are employed during the acquisition process, thereby effectively reducing the impact of environmental noise and interference signals. The acquisition duration of the raw vibration signal is determined by the signal monitoring duration, which is typically set according to actual operating conditions to ensure the integrity and representativeness of the acquired data.

[0021] After acquiring the original vibration signal, preprocessing is required. The main purposes of preprocessing are to remove trend terms from the signal, normalize the signal amplitude range, and filter out high-frequency noise. Detrending is achieved through polynomial fitting or the difference method, while normalization uses the min-max normalization method to unify the signal amplitude within the [0,1] interval. A bandpass filter is selected for filtering, with its cutoff frequency set according to the target frequency range to retain frequency components relevant to the fault characteristics. After the above preprocessing, the initial vibration signal is obtained, laying the foundation for subsequent analysis.

[0022] Next, broadband mode decomposition (BMD) is performed on the initial vibration signal. BMD is an efficient signal decomposition method that can decompose complex signals into multiple modal components with different frequency characteristics. Specifically, the decomposed modal components are first segmented according to a preset frequency range to obtain low-frequency, mid-frequency, and high-frequency modal components. Then, a sliding time window technique is used to further segment the modal components in each frequency band, generating multiple sets of time-window low-frequency, mid-frequency, and high-frequency modal sequences. Amplitude peaks are extracted sequentially from each time-window sequence to form multiple sets of low-frequency, mid-frequency, and high-frequency energy point sequences. Further, nonlinear fitting is performed on these energy point sequences to obtain the target low-frequency, mid-frequency, and high-frequency energy distributions, respectively. Finally, the three energy distribution matrices are concatenated to generate the target energy distribution matrix. This process is as follows: Fig. 2 As shown, each decomposition step and energy extraction step is clearly labeled.

[0023] After obtaining the target energy distribution matrix, the bearing fault mode determination stage begins. First, bearing fault modes are extracted sequentially from a pre-defined set of bearing fault modes, and their typical energy distribution matrices are obtained. The generation process of the typical energy distribution matrix includes the following steps: multiple energy distribution matrices are extracted from the set of energy distribution matrices for bearing fault modes, and these matrices are trimmed according to the signal monitoring duration and frequency band input by the user to obtain a set of energy distribution matrices to be averaged. Next, the signal monitoring time points and frequency points are determined according to the signal monitoring duration and frequency band, constructing an energy calibration point matrix. Energy calibration points are extracted sequentially from the energy calibration point matrix, and the corresponding energy value sets to be averaged are identified in the set of energy distribution matrices to be averaged. The average energy value of the energy value set to be averaged is calculated, and three-dimensional energy points are generated by combining the energy calibration points. The three-dimensional energy points of all energy calibration points are summarized to form a three-dimensional energy point set, and then a typical energy distribution matrix is ​​generated using a surface fitting method.

[0024] After obtaining the typical energy distribution matrix, the first degree of matching between the target energy distribution matrix and the typical energy distribution matrix is ​​calculated according to a preset energy matching rule. The core formula of the energy matching rule is as follows: ; in, The first degree of matching between the typical energy distribution matrix of the i-th bearing failure mode and the target energy distribution matrix is ​​represented, and N represents the number of energy calibration points in the energy calibration matrix. This represents the energy value of the p-th energy calibration point in the target energy distribution matrix. Let represent the energy value at the p-th energy calibration point in the typical energy distribution matrix of the i-th bearing failure mode. The absolute value sign is used to measure the degree of difference between the two. A first matching degree set is formed by calculating the first matching degree of all bearing failure modes.

[0025] The second determination method involves the introduction of reference energy fluctuation interval sets. The user inputs multiple sets of reference energy fluctuation interval sets based on the bearing fault mode. Each set includes the reference waveform period, reference waveform amplitude, and reference waveform steepness. Simultaneously, multiple sets of target energy fluctuation interval sets are identified from the target energy distribution matrix. Specifically, signal monitoring frequencies are extracted sequentially from the signal monitoring frequency points, and the energy change curves corresponding to those frequencies in the target energy distribution matrix are identified. By analyzing the target waveform period set, target waveform amplitude set, and target waveform steepness set of the energy change curves, the target average waveform period, target average waveform amplitude, and target average waveform steepness are calculated respectively, forming the target energy fluctuation interval sets. The target energy fluctuation interval sets for each signal monitoring frequency are then aggregated to obtain multiple sets of target energy fluctuation interval sets.

[0026] Subsequently, the second degree of matching between multiple sets of target energy fluctuation intervals and multiple sets of reference energy fluctuation intervals is calculated. The formula for calculating the second degree of matching is as follows: ; in, denoted by , represents the second matching degree between multiple sets of target energy fluctuation intervals and multiple sets of reference energy fluctuation intervals for the i-th bearing failure mode; Q represents the number of signal monitoring frequencies among the signal monitoring frequency points. , and These represent the period coefficient, amplitude coefficient, and kurtosis coefficient, respectively. and Let these represent the reference waveform period and the target average waveform period, respectively, within the q-th set of reference energy fluctuation intervals. and Let these represent the reference waveform amplitude and the target average waveform amplitude, respectively, within the q-th reference energy fluctuation interval set. and Let represent the steepness of the reference waveform and the steepness of the target average waveform in the q-th reference energy fluctuation interval set, respectively. A second matching degree set is formed by calculating the second matching degree for all bearing failure modes.

[0027] The third determination method utilizes a target deep learning model for probabilistic prediction. First, based on a preset amplitude sampling interval, amplitude sequences are extracted from the energy variation curves of the signal monitoring frequencies, forming amplitude sequence sets corresponding to different signal monitoring frequencies. Then, these amplitude sequence sets are input into the target deep learning model, which employs a convolutional neural network structure. Its input layer includes multiple amplitude sequence input nodes, and its output layer corresponds to multiple fault mode probability value output nodes. Through model prediction, a third matching probability value set is obtained.

[0028] After completing the above three determination methods, the target matching confidence set needs to be calculated using a comprehensive matching formula based on the preset multi-dimensional weight set, the first matching degree set, the second matching degree set, and the third matching probability value set. The specific form of the comprehensive matching formula is as follows: ; in, The target energy distribution matrix represents the target matching confidence level for the i-th bearing failure mode. , and These represent the first fusion weight, the second fusion weight, and the third fusion weight, respectively. This represents the degree of matching between the target energy distribution matrix in the first matching set and the typical energy distribution matrix of the i-th bearing failure mode. This represents the matching degree between the target energy fluctuation interval set in the second matching degree set and the energy fluctuation interval set of the i-th bearing failure mode. Let represent the probability value of the failure mode of the target energy distribution matrix in the third matching probability value set as the i-th bearing failure mode. A target matching confidence set is formed by calculating the target matching confidence of all bearing failure modes.

[0029] The process of obtaining the multi-dimensional weight set involves iterative optimization of the training data. Specifically, training energy distribution matrix sets are extracted sequentially from multiple pre-defined sets of training energy distribution matrices, and training energy distribution matrices are extracted sequentially from each training set. A first training matching degree set of the training energy distribution matrix is ​​calculated based on typical energy distribution matrices and energy matching rules, and the first predicted fault mode corresponding to the maximum first training matching degree is identified, forming a first predicted fault mode set. Simultaneously, multiple sets of training energy fluctuation intervals of the training energy distribution matrix are identified, and a second training matching degree set is calculated based on multiple sets of reference energy fluctuation intervals and multiple sets of training energy fluctuation intervals, and the second predicted fault mode corresponding to the maximum second training matching degree is identified, forming a second predicted fault mode set. Furthermore, a third set of training probability values ​​of the training energy distribution matrix is ​​predicted using the target deep learning model, and the third predicted fault mode corresponding to the maximum third training probability value is identified, forming a third predicted fault mode set. The actual fault modes of the training energy distribution matrix are obtained, forming an actual fault mode set. The sequence consistency between the first, second, and third predicted fault mode sets and the actual fault mode set is calculated respectively, yielding first, second, and third sequence consistency. Determine whether the consistency of these sequences tends to be stable. If it does not tend to be stable, return to the training data extraction step. If it tends to be stable, normalize the consistency of the first sequence, the consistency of the second sequence, and the consistency of the third sequence to obtain a multi-dimensional weight set.

[0030] The process of determining the stability of sequence consistency includes the following steps: obtaining a first sequence-consistent sequence, a second sequence-consistent sequence, and a third sequence-consistent sequence, and plotting the first sequence consistency curve, the second sequence consistency curve, and the third sequence consistency curve, respectively. Identifying the first, second, and third terminal slopes of these curves, and determining whether these slopes are all less than a preset terminal slope threshold. If the condition is not met, the sequence consistency has not stabilized; if the condition is met, the sequence consistency has stabilized.

[0031] Finally, the maximum target matching confidence score is identified in the target matching confidence score set, and the target bearing failure mode corresponding to this maximum target matching confidence score is identified. This process is completed by the target bearing failure mode identification module, whose functional architecture is as follows: Fig. 3As shown. Through the above steps, this invention achieves accurate diagnosis of fault modes in low-speed, heavy-load bearings, significantly improving the accuracy, robustness, and real-time performance of fault diagnosis.

[0032] In summary, this invention solves the problem of fault diagnosis of low-speed heavy-load bearings under complex working conditions through a series of rigorous technical means and algorithm design, providing strong support for health monitoring and maintenance of industrial equipment.

Claims

1. A method for mining fault features of low-speed, heavy-load bearings based on broadband mode decomposition, characterized in that, The method includes: The original vibration signal of the bearing is acquired by a pre-constructed multi-channel vibration sensor. The acquisition process of the original vibration signal includes: vibration signal enhancement, noise suppression and blind source separation. The duration of the original vibration signal is the signal monitoring duration. The original vibration signal is preprocessed to obtain an initial vibration signal, wherein the preprocessing includes: detrending, normalization and filtering; The initial vibration signal is subjected to broadband mode decomposition to obtain multiple broadband mode components. The energy distribution of the multiple broadband mode components is extracted according to a preset frequency range to obtain the target energy distribution matrix. Bearing failure modes are extracted sequentially from a preset set of bearing failure modes, and a typical energy distribution matrix of the bearing failure mode is obtained. The first matching degree between the target energy distribution matrix and the typical energy distribution matrix is ​​calculated according to a preset energy matching rule to obtain a first matching degree set. Receive multiple sets of reference energy fluctuation intervals input by the user according to the bearing failure mode, identify multiple sets of target energy fluctuation intervals of the target energy distribution matrix, calculate the second matching degree between the multiple sets of target energy fluctuation intervals and the multiple sets of reference energy fluctuation intervals, and obtain the second matching degree set, wherein the reference energy fluctuation interval set includes: reference waveform period, reference waveform amplitude and reference waveform steepness; The target energy distribution matrix is ​​input into a pre-constructed target deep learning model to obtain a third matching probability value set, wherein the target deep learning model is a convolutional neural network; Based on the preset multi-dimensional weight set, the first matching degree set, the second matching degree set, and the third matching probability value set, the target matching confidence set is calculated using a pre-constructed comprehensive matching formula; Identify the maximum target matching confidence score in the target matching confidence score set, and identify the target bearing failure mode corresponding to the maximum target matching confidence score.

2. The method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition as described in claim 1, characterized in that, The step of extracting the energy distribution of the multiple sets of broadband mode components according to a preset frequency range to obtain a target energy distribution matrix includes: The multiple broadband modal components are divided once according to the frequency range to obtain low-frequency modal components, mid-frequency modal components and high-frequency modal components; The low-frequency mode component, mid-frequency mode component, and high-frequency mode component are further segmented according to a preset time window to obtain multiple sets of time window low-frequency mode sequences, multiple sets of time window mid-frequency mode sequences, and multiple sets of time window high-frequency mode sequences. Amplitude peak values ​​are extracted sequentially from the multiple sets of low-frequency mode sequences, multiple sets of mid-frequency mode sequences, and multiple sets of high-frequency mode sequences within time windows to obtain multiple sets of low-frequency energy point sequences, multiple sets of mid-frequency energy point sequences, and multiple sets of high-frequency energy point sequences. Energy fitting is performed on the multiple sets of low-frequency energy point sequences, multiple sets of mid-frequency energy point sequences, and multiple sets of high-frequency energy point sequences respectively to obtain the target low-frequency energy distribution, the target mid-frequency energy distribution, and the target high-frequency energy distribution. The target low-frequency energy distribution, target mid-frequency energy distribution, and target high-frequency energy distribution are matrix-stitched to obtain the target energy distribution matrix.

3. The method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition as described in claim 1, characterized in that, The acquisition of the typical energy distribution matrix of bearing failure modes includes: Obtain the energy distribution matrix set of the bearing failure mode; Energy distribution matrices are extracted sequentially from the energy distribution matrix set. The user then clips the energy distribution matrix according to the signal monitoring duration and signal monitoring frequency band to obtain the set of energy distribution matrices to be averaged. The signal monitoring time point and signal monitoring frequency point are determined based on the signal monitoring duration and signal monitoring frequency range, and an energy calibration point array is constructed based on the signal monitoring time point and signal monitoring frequency point; Energy calibration points are extracted sequentially from the energy calibration point matrix, and the corresponding set of energy values ​​to be averaged is identified in the set of energy distribution matrix to be averaged based on the energy calibration points. Calculate the average energy value of the set of energy values ​​to be averaged, and determine the three-dimensional energy points based on the average energy value and the energy calibration points; By summing the three-dimensional energy points of each energy calibration point, a three-dimensional energy point set is obtained; By fitting the three-dimensional energy point set, a typical energy distribution matrix is ​​obtained.

4. The method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition as described in claim 1, characterized in that, The energy matching rules are as follows: ； in, The first degree of matching between the typical energy distribution matrix of the i-th bearing failure mode and the target energy distribution matrix is ​​represented, and N represents the number of energy calibration points in the energy calibration matrix. This represents the energy value of the p-th energy calibration point in the target energy distribution matrix. This represents the energy value at the p-th energy calibration point in the typical energy distribution matrix for the i-th bearing failure mode.

5. The method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition as described in claim 1, characterized in that, The set of multiple target energy fluctuation intervals for identifying the target energy distribution matrix includes: The signal monitoring frequencies are extracted sequentially from the signal monitoring frequency points, and the energy change curves of the signal monitoring frequencies are identified in the target energy distribution matrix. Identify the target waveform period set, target waveform amplitude set, and target waveform steepness set of the energy change curve; Based on the target waveform period set, target waveform amplitude set, and target waveform steepness set, the target average waveform period, target average waveform amplitude, and target average waveform steepness are calculated respectively to obtain the target energy fluctuation interval set. By aggregating the target energy fluctuation interval sets of various signal monitoring frequencies, multiple sets of target energy fluctuation interval sets are obtained.

6. The method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition as described in claim 1, characterized in that, The calculation of the second matching degree between the multiple sets of target energy fluctuation intervals and the multiple sets of reference energy fluctuation intervals yields a second matching degree set, including: The second matching degree is calculated based on the multiple sets of target energy fluctuation intervals and the multiple sets of reference energy fluctuation intervals: ； in, denoted by , represents the second matching degree between multiple sets of target energy fluctuation intervals and multiple sets of reference energy fluctuation intervals for the i-th bearing failure mode; Q represents the number of signal monitoring frequencies among the signal monitoring frequency points. , and These represent the period coefficient, amplitude coefficient, and kurtosis coefficient, respectively. and Let these represent the reference waveform period and the target average waveform period, respectively, within the q-th set of reference energy fluctuation intervals. and Let these represent the reference waveform amplitude and the target average waveform amplitude, respectively, within the q-th reference energy fluctuation interval set. and These represent the steepness of the reference waveform and the steepness of the target average waveform within the q-th group of reference energy fluctuation intervals, respectively.

7. The method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition as described in claim 1, characterized in that, The step of inputting the target energy distribution matrix into a pre-constructed target deep learning model to obtain a third matching probability value set includes: Based on a preset amplitude sampling interval, the amplitude sequence is extracted from the energy change curve of the signal monitoring frequency to obtain the amplitude sequence set corresponding to different signal monitoring frequencies; The amplitude sequence set is input into the target deep learning model to obtain a third matching probability value set. The input layer of the target deep learning model includes: a first amplitude sequence input node, a second amplitude sequence input node, and an nth amplitude sequence input node. The output layer of the target deep learning model includes: a first fault mode probability value output node, a second fault mode probability value output node, and an nth fault mode probability value output node.

8. The method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition as described in claim 1, characterized in that, The methods for obtaining the multi-dimensional weight set include: Extract training energy distribution matrix sets sequentially from multiple pre-set training energy distribution matrix sets; The training energy distribution matrix is ​​extracted sequentially from the set of training energy distribution matrices. Calculate the first training matching degree set of the training energy distribution matrix according to the typical energy distribution matrix and the energy matching rule, identify the maximum first training matching degree in the first training matching degree set, identify the first predicted fault mode corresponding to the maximum first training matching degree, and obtain the first predicted fault mode set. Identify multiple sets of training energy fluctuation intervals of the training energy distribution matrix, calculate a second training matching degree set based on the multiple sets of reference energy fluctuation intervals and the multiple sets of training energy fluctuation intervals, identify the maximum second training matching degree in the second training matching degree set, identify the second predicted fault mode corresponding to the maximum second training matching degree, and obtain a second predicted fault mode set. The target deep learning model is used to predict the third training probability value set of the training energy distribution matrix. The maximum third training probability value is identified in the third training probability value set, and the third predicted fault mode corresponding to the maximum third training probability value is identified to obtain the third predicted fault mode set. Obtain the true fault modes from the training energy distribution matrix to obtain the true fault mode set; The sequence consistency between the first predicted fault mode set, the second predicted fault mode set, and the third predicted fault mode set and the actual fault mode set is calculated respectively to obtain the first sequence consistency, the second sequence consistency, and the third sequence consistency. Determine whether the consistency of the first sequence, the consistency of the second sequence, and the consistency of the third sequence tend to be stable; if they do not tend to be stable, return to the steps of sequentially extracting the training energy distribution matrix set from the preset set of multiple training energy distribution matrices; if they tend to be stable, normalize the consistency of the first sequence, the consistency of the second sequence, and the consistency of the third sequence to obtain a multi-dimensional weight set.

9. The method for mining fault features of low-speed heavy-load bearings based on broadband mode decomposition as described in claim 8, characterized in that, The determination of whether the consistency of the first sequence, the consistency of the second sequence, and the consistency of the third sequence tend to be stable includes: Obtain the first sequence-consistent sequence, the second sequence-consistent sequence, and the third sequence-consistent sequence; Plot the first sequence consensus curve, the second sequence consensus curve, and the third sequence consensus curve based on the first sequence consensus curve, the second sequence consensus curve, and the third sequence consensus curve, respectively. Identify the first terminal slope, second terminal slope, and third terminal slope of the first sequence consistency curve, the second sequence consistency curve, and the third sequence consistency curve; Determine whether the first terminal slope, the second terminal slope, and the third terminal slope are all less than a preset terminal slope threshold; if the condition is not met, the first sequence consistency, the second sequence consistency, and the third sequence consistency have not stabilized; if the condition is met, the first sequence consistency, the second sequence consistency, and the third sequence consistency have stabilized.

10. A system for mining fault features of low-speed, heavy-load bearings based on broadband mode decomposition, characterized in that, The system includes: The target energy distribution matrix acquisition module is used to acquire the original vibration signal of the bearing through a pre-constructed multi-channel vibration sensor. The acquisition process of the original vibration signal includes: vibration signal enhancement, noise suppression, and blind source separation. The duration of the original vibration signal is the signal monitoring duration. The original vibration signal is preprocessed to obtain an initial vibration signal. The preprocessing includes: detrending, normalization, and filtering. The initial vibration signal is subjected to broadband mode decomposition to obtain multiple broadband mode components. The energy distribution of the multiple broadband mode components is extracted according to a preset frequency range to obtain the target energy distribution matrix. The matching degree calculation module is used to sequentially extract bearing failure modes from a preset bearing failure mode set, obtain a typical energy distribution matrix of the bearing failure mode, calculate the first matching degree between the target energy distribution matrix and the typical energy distribution matrix according to a preset energy matching rule, and obtain a first matching degree set; receive multiple sets of reference energy fluctuation intervals input by the user according to the bearing failure mode, identify multiple sets of target energy fluctuation intervals of the target energy distribution matrix, calculate the second matching degree between the multiple sets of target energy fluctuation intervals and the multiple sets of reference energy fluctuation intervals, and obtain a second matching degree set, wherein the reference energy fluctuation intervals include: reference waveform period, reference waveform amplitude, and reference waveform steepness; input the target energy distribution matrix into a pre-constructed target deep learning model to obtain a third matching probability value set, wherein the target deep learning model is a convolutional neural network; The target matching confidence set calculation module is used to calculate the target matching confidence set based on the preset multi-dimensional weight set, the first matching degree set, the second matching degree set and the third matching probability value set, using a pre-constructed comprehensive matching formula. The target bearing failure mode identification module is used to identify the maximum target matching confidence score in the target matching confidence score set, and to identify the target bearing failure mode corresponding to the maximum target matching confidence score.

Images