Rolling bearing fault diagnosis method and system based on deep neural network

Abstract

The invention provides a rolling bearing fault diagnosis method and system based on a deep neural network, and relates to the technical field of mechanical intelligent manufacturing. According to the method, the original vibration signal can be preprocessed to obtain the vibration signal to be analyzed, modal decomposition is carried out based on variational modal decomposition, in the process, the working condition change influence index is generated based on working condition change data, whether the working condition change influence index is higher than the working condition influence degree threshold value or not is judged, and if yes, the vibration signal to be analyzed is obtained. Secondly, judging whether the signal-to-noise ratio of the original vibration signal is lower than the minimum value of the signal-to-noise ratio or not, if yes, changing the first-class adjustment into second-class adjustment, if not, carrying out predictive adjustment judgment, and after modal decomposition is completed, carrying out second-class adjustment; fault diagnosis is carried out based on the deep neural network, the accuracy of early fault diagnosis of the rolling bearing is improved, and the problem that the early fault diagnosis capability of the rolling bearing is insufficient in the prior art is solved.

Description

A method and system for fault diagnosis of rolling bearings based on deep neural networks Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method and system for diagnosing rolling bearing faults based on deep neural networks. Background Technology

[0002] Rolling bearings are among the most commonly used and critical basic components in rotating machinery, and their operating condition directly determines the operational safety and lifespan of the equipment. Bearings are widely used in motors, fans, CNC machine tools, rail vehicles, and aero engines. Once a bearing fails, it can lead to degradation of overall machine performance, equipment downtime, or even safety accidents. Therefore, accurate and timely diagnosis of early rolling bearing failures is an important technical means to ensure the reliable operation of mechanical systems.

[0003] As a critical load-bearing component in rotating machinery, rolling bearings are difficult to accurately identify using traditional methods. Deep neural networks, with their end-to-end feature learning capabilities and powerful nonlinear expression capabilities, offer a solution. Researching rolling bearing fault diagnosis methods based on deep neural networks can not only overcome the limitations of traditional feature engineering and achieve adaptive extraction and identification of weak signal features, but also promote the development of intelligent operation and maintenance technology for machinery. This has significant theoretical research value and broad engineering application value.

[0004] For example, Chinese invention patent CN120800803A discloses a method and apparatus for diagnosing rolling bearing faults based on a deep convolutional neural network, which includes: acquiring vibration signals of the rolling bearing under multiple working conditions, adding labels and dividing them into training and testing sets; using the training and testing sets to train and diagnose the deep convolutional neural network to obtain a diagnostic model; the deep convolutional neural network includes multiple network units connected in sequence, including a convolutional module, an efficient temporal channel attention module and a max pooling layer connected in sequence; the efficient temporal channel attention module takes the output of the convolutional module as input, adopts parallel channel attention modules and temporal attention modules, adaptively allocates attention weights in each time segment and channel dimension, and performs weighted fusion of the output attention map and input features to filter out important feature information and perform dimensionality reduction.

[0005] For example, the Chinese invention patent with announcement number CN112903294B discloses a method for diagnosing rolling bearing faults based on variational mode decomposition and deep convolutional neural networks, which includes: Step 1: collecting raw vibration data of the rolling bearing; Step 2: performing variational mode decomposition data processing and neural network training on the training set vibration data; Step 3: performing variational mode decomposition data processing on the test vibration data and using a neural network for fault diagnosis.

[0006] Existing technologies for fault diagnosis of rolling bearings based on deep neural networks essentially utilize deep learning models to automatically extract features from bearing vibration signals and classify faults. The specific implementation process generally includes data acquisition and preprocessing, data conversion and enhancement, deep neural network feature extraction and classification, and outputting fault categories. Among these, data acquisition and preprocessing is used to collect vibration acceleration or velocity signals and perform preprocessing such as denoising and normalization; data conversion and enhancement is used to convert one-dimensional signals into time-frequency images (such as short-time Fourier transform, Gramian angle field, etc.) or to perform data enhancement; deep neural network feature extraction and classification is used to learn discriminative features directly from the data using convolutional neural networks, residual networks, attention mechanism networks, etc.; and outputting fault categories is used to output the probability of different fault types through the last layer classifier (such as Softmax).

[0007] The above-mentioned technology has at least the following technical problems:

[0008] Because early fault signals have extremely low energy, non-stationary characteristics, and are easily affected by noise, the vibration signals generated by early rolling bearing faults (such as microcracks and early pitting) are extremely weak, have a low signal-to-noise ratio, and their characteristic frequencies are easily submerged by noise. Deep networks often struggle to capture these minute changes in the data, resulting in insufficient ability to diagnose early rolling bearing faults and making it difficult to achieve true predictive maintenance. Summary of the Invention

[0009] Therefore, embodiments of the present invention provide a method and system for diagnosing rolling bearing faults based on deep neural networks, which can improve the accuracy of early fault diagnosis of rolling bearings.

[0010] The technical solution of this invention is implemented as follows: Within a current preset time window, the original vibration signal of the rolling bearing, its signal-to-noise ratio, and operating condition change data are acquired. The operating condition change data reflects the changes in the rolling bearing as the actual working environment changes. The original vibration signal is preprocessed to obtain the corresponding vibration signal to be analyzed. Modal decomposition is performed on the vibration signal to be analyzed based on the decomposition parameters in variational mode decomposition, including the number of decomposition levels and penalty parameters. During the variational mode decomposition process, an operating condition change influence index is generated based on the operating condition change data. It is determined whether the operating condition change influence index is higher than a preset operating condition influence threshold. If so, the difference between the operating condition change influence index and the operating condition influence threshold is used to determine the influence index. The adaptive adjustment of the decomposition parameters is performed in a first-class manner; otherwise, variational mode decomposition continues. It is determined whether the signal-to-noise ratio (SNR) of the original vibration signal is lower than the minimum SNR. If so, the first-class adjustment of the adaptive adjustment of the decomposition parameters is changed to a second-class adjustment based on the SNR. Otherwise, a predictive adjustment is performed based on the SNR and the operating condition change influence index to determine whether to perform predictive adjustment of the decomposition parameters for the next preset time window. After mode decomposition is completed, the decomposed modal components are output, and the enhanced vibration signal is reconstructed based on the modal components. This signal is then input into a deep neural network for fault diagnosis, and the reliability of the fault diagnosis results is evaluated to determine whether to output the operating condition data sampling feedback command for the next preset time window.

[0011] This invention also provides a rolling bearing fault diagnosis system based on a deep neural network. The method includes: a data sampling module: acquiring the original vibration signal of the rolling bearing, its signal-to-noise ratio, and operating condition change data within a current preset time window, wherein the operating condition change data reflects the changes in the rolling bearing with changes in the actual working environment; a data processing and decomposition module: preprocessing the original vibration signal to obtain the corresponding vibration signal to be analyzed, and performing modal decomposition on the vibration signal to be analyzed based on the decomposition parameters in variational mode decomposition, including the number of decomposition layers and a penalty parameter; and an operating condition influence decomposition module: generating an operating condition change influence index based on the operating condition change data during the variational mode decomposition process, and determining whether the operating condition change influence index is higher than a preset operating condition influence degree threshold. If so, the operating condition change influence index is compared with the operating condition influence degree threshold. The difference between the threshold values ​​of the influence of the operating conditions affects the adaptive adjustment of the decomposition parameters. If the difference is not met, the adaptive adjustment of the decomposition parameters is adjusted accordingly. Otherwise, variational mode decomposition continues. The signal-to-noise ratio (SNR) adjustment module determines whether the SNR of the original vibration signal is lower than the minimum SNR. If so, it changes the adaptive adjustment of the decomposition parameters from the first-class adjustment to the second-class adjustment based on the SNR. Otherwise, it performs predictive adjustment based on the SNR and the operating condition change influence index to determine whether to perform predictive adjustment of the decomposition parameters for the next preset time window. The fault diagnosis result output module outputs the decomposed modal components after mode decomposition is completed. Based on these modal components, it reconstructs the signal to obtain an enhanced vibration signal, which is then input into a deep neural network for fault diagnosis. The reliability of the fault diagnosis results is assessed to determine whether to output the operating condition data sampling feedback command for the next preset time window.

[0012] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0013] 1. Within the current preset time window, the original vibration signal is preprocessed to obtain the corresponding vibration signal to be analyzed. Modal decomposition is then performed on the vibration signal based on the decomposition parameters in variational mode decomposition, which helps improve the decomposition accuracy and stability. During variational mode decomposition, it is determined whether to perform a type I adjustment to the adaptive adjustment of the decomposition parameters based on changes in operating conditions; otherwise, variational mode decomposition continues. This achieves automatic optimization of decomposition characteristics when operating conditions change significantly, avoiding mode aliasing and helping to maintain stable decomposition results. Next, it is determined whether the signal-to-noise ratio (SNR) of the original vibration signal is lower than the minimum SNR. If so, the type I adjustment to the adaptive adjustment of the decomposition parameters is changed to a type II adjustment based on the SNR. This not only helps enhance the energy performance of the fault impact component but also effectively suppresses background noise. Otherwise, the adjustment is based on the SNR and changes in operating conditions. The predictive adjustment of the influence index determines whether to predictively adjust the decomposition parameters for the next preset time window. This helps to optimize the algorithm's computational resources and decomposition convergence speed in advance, enabling real-time processing of continuous data streams. After the modal decomposition is completed, the decomposed modal components are output, and the enhanced vibration signal is obtained by reconstructing the signal based on the modal components. This improves the model's ability to perceive weak fault features under different working conditions. The signal is then input into a deep neural network for fault diagnosis, and the reliability of the fault diagnosis results is evaluated to determine whether to output the working condition data sampling feedback command for the next preset time window. This not only achieves closed-loop adaptive optimization of the diagnosis results, data acquisition, and decomposition parameters, but also achieves high-precision, robust decomposition and feature enhancement of rolling bearing vibration signals, ensuring the accuracy of rolling bearing fault identification.

[0014] 2. By quantifying the difference between the operating condition change influence index and the operating condition influence degree threshold, the operating condition change difference value is obtained. This numerically reflects the intensity and direction of the operating condition disturbance, accurately characterizes the actual operating condition change amplitude, and provides a quantifiable basis for parameter adjustment. Then, the operating condition change difference value is input into the pre-trained decomposition layer influence sequence and penalty parameter influence sequence, respectively, and the corresponding adjustment coefficient influence amount is output, realizing the mapping learning of parameter adjustment and improving the adaptability of mode decomposition. Then, by multiplying the adjustment coefficient influence amount with the initial adjustment coefficient in variational mode decomposition, a class of adjustment coefficients is obtained. This mechanism not only realizes the dynamic adaptive update of parameters, but also ensures that the decomposition process responds to operating condition changes in real time. Finally, the initial adjustment coefficient used to obtain the decomposition parameters is replaced with a class of adjustment coefficients to complete the variational mode decomposition and output the corresponding mode components. This realizes the universality, scalability, and long-term adaptability of the decomposition parameter adjustment of variational mode decomposition, which helps to improve the identification accuracy of early fault signals of rolling bearings.

[0015] 3. By performing a difference calculation between the signal-to-noise ratio (SNR) and the minimum SNR of the original vibration signal, and then normalizing the data to obtain the SNR offset, it is helpful to extract the early weak fault impact components. Then, a weighted fusion calculation is performed on the SNR offset and the difference value of the operating condition changes to obtain the fusion influence of the decomposition parameters. This allows the decomposition parameters to simultaneously respond in a coordinated manner to external operating condition disturbances and internal signal quality changes. The fusion influence of the decomposition parameters is then compared in the preset decomposition parameter mapping table and the penalty parameter mapping table to obtain the corresponding second-order adjustment coefficients. Through this mechanism, variational mode decomposition can maintain the frequency concentration and energy consistency of modal components even under strong noise and complex operating conditions. Finally, the first-order adjustment coefficients used for the decomposition parameters are replaced with second-order adjustment coefficients to complete the variational mode decomposition and output the corresponding modal components. This not only ensures dynamic and robust decomposition of roller fault signals under multiple noise and operating conditions, but also provides technical support for higher robustness and reliability in intelligent fault diagnosis of rolling bearings. Attached Figure Description

[0016] Figure 1 is a flowchart illustrating the rolling bearing fault diagnosis method based on deep neural networks provided in an embodiment of the present invention.

[0017] Figure 2 is a schematic diagram of the predictive adjustment process provided in an embodiment of the present invention;

[0018] Figure 3 is a schematic diagram of the structure of the rolling bearing fault diagnosis system based on deep neural network provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0020] In this application, the terms "first," "second," "third," etc., are used to distinguish identical or similar items with substantially the same function and purpose. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor does it limit the quantity or execution order. It should also be understood that although the following description uses the terms "first," "second," etc., to describe various elements, these elements should not be limited by the terms. These terms are merely used to distinguish one element from another. For example, "first device," "second device," "third device," etc., are only used to distinguish devices. Similarly, "first sample data," "second sample data," and "third sample data," etc., are only used to distinguish sample data. Without departing from the scope of the various examples, a first device can be referred to as a second device, and similarly, a second device can be referred to as a first device. Both the first device and the second device are devices, and in some cases, they can be separate and distinct devices.

[0021] Example 1:

[0022] This invention provides a method and system for diagnosing rolling bearing faults based on deep neural networks. Figure 1 shows a flowchart of the rolling bearing fault diagnosis method based on deep neural networks provided in this invention. The processing flow of this method may include the following steps:

[0023] Step 1: Within the current preset time window, acquire the original vibration signal of the rolling bearing, its signal-to-noise ratio, and the operating condition change data. The operating condition change data reflects the changes in the rolling bearing as the actual working environment changes.

[0024] It should be added that the vibration signal of the rolling bearing is collected by an accelerometer, that is, the original vibration signal; the signal-to-noise ratio (SNR) essentially measures the ratio of "signal component energy" to "noise energy". The noise component is separated in the original vibration signal by time-frequency domain filtering methods (such as bandpass filtering, wavelet threshold denoising, etc.), and then the corresponding SNR is calculated by the definition of SNR.

[0025] Step 2: Preprocess the original vibration signal to obtain the corresponding vibration signal to be analyzed, and perform modal decomposition on the vibration signal to be analyzed based on the decomposition parameters in variational mode decomposition. The decomposition parameters include the number of decomposition layers and the penalty parameter.

[0026] It should be explained that the methods for preprocessing the original vibration signal include, but are not limited to, filtering and denoising, detrending and normalization, envelope analysis, signal enhancement and segmented sampling; variational mode decomposition is an adaptive signal decomposition algorithm based on the variational principle. It decomposes the original non-stationary signal into several physically meaningful narrowband modes through bandwidth minimization and frequency separation constraints, and has high stability, strong noise resistance and excellent frequency resolution.

[0027] Wherein, the number of decomposition layers represents the number of modal components to be decomposed, and its expression is: The penalty parameter controls the smoothness of the bandwidth for each mode, and its expression is: , Indicates the number of decomposition layers (or the number of modes). Indicates the penalty parameter. This indicates the initial number of decomposition layers, the base number set under standard signal-to-noise ratio or default conditions. This represents the initial penalty parameter, the initial penalty factor set under standard signal-to-noise ratio conditions. This indicates an initial adjustment coefficient of 1, controlling the effect of signal-to-noise ratio changes on the number of decomposition layers. The intensity of the impact, This represents an initial adjustment coefficient of 2, which controls the impact of changes in the signal-to-noise ratio on the penalty parameter. The intensity of the impact, This represents a monotonically decreasing function that characterizes the relationship between the signal-to-noise ratio and the adjustment amplitude. Signal-to-noise ratio (SNR) represents the ratio of effective information energy to noise energy in a signal, reflecting signal quality.

[0028] Step 3: During variational mode decomposition, an operating condition change influence index is generated based on the operating condition change data. It is then determined whether the operating condition change influence index is higher than the preset operating condition influence threshold. If so, the adaptive adjustment of the decomposition parameters is adjusted based on the difference between the operating condition change influence index and the operating condition influence threshold. Otherwise, variational mode decomposition continues.

[0029] It should be added that the threshold for the degree of influence of operating conditions is a data that limits the maximum impact of changes in operating conditions on rolling bearings. It is generally preset by the staff based on historical data and experience rules within a historical time period and stored in a preset database. When needed, it can be directly extracted from the preset database.

[0030] It's important to explain that before designing a deep neural network-based rolling bearing fault diagnosis method, technical professionals typically pre-build a preset database to support various control strategies. This database integrates multiple key control parameters, including preset time windows, thresholds for the degree of influence of operating conditions, decomposition parameter mapping tables, penalty parameter mapping tables, preset differences, first preset control amplitude, second preset control amplitude, first prediction interval, second prediction interval, weighting coefficients, and reference data. For the analysis logic and actual hardware environment of the deep neural network-based rolling bearing fault diagnosis method, technical professionals have pre-set all parameters. This resulting preset database provides crucial data support for subsequent data uploading, storage optimization, and even automated screening and judgment.

[0031] The specific methods for obtaining the index of the impact of changes in operating conditions are as follows:

[0032] The first step is to normalize the operating condition change data to obtain standard operating condition change data, which includes speed, load torque, motor current, motor voltage, and bearing load.

[0033] Specifically, the speed is output by the frequency converter or control system, the load torque can be measured by a torque sensor, the motor current is measured by a Hall current sensor, the motor voltage is measured by a voltage sensor, and the bearing load is directly measured by a strain gauge force sensor.

[0034] The second step is to smooth the standard operating condition variation data to obtain the operating condition variation impact index. The operating condition variation difference value is used to quantify the degree of influence of the actual operating conditions of the rolling bearing on the variational mode decomposition. The aforementioned smoothing of the standard operating condition data specifically means averaging the data of each standard operating condition.

[0035] One type of adjustment involves adaptively regulating the decomposition parameters based on the difference between the operating condition change influence index and the operating condition influence degree threshold. The specific process is as follows:

[0036] First, the difference between the operating condition change impact index and the operating condition impact degree threshold is quantified to obtain the operating condition change difference value. That is, the difference between the operating condition change impact index and the operating condition impact degree threshold is calculated, and then the ratio of the difference between the operating condition change impact index and the operating condition impact degree threshold is calculated to obtain the operating condition change difference value.

[0037] Next, the difference values ​​of the working condition changes are input into the pre-trained decomposition layer number influence sequence and the penalty parameter influence sequence, respectively, and the corresponding adjustment coefficient influence amount, namely the decomposition layer number influence amount and the penalty parameter influence amount are output respectively.

[0038] Specifically, by inputting the operating condition variation difference values ​​into the decomposition layer influence sequence and the penalty parameter influence sequence, the corresponding adjustment coefficient influence can be obtained. The aforementioned data sequence is used to fit the mapping relationship between the operating condition variation difference values ​​and the adjustment coefficient influence. The construction method is as follows: in the initial data sequence constructed based on the gradient boosting regression algorithm, the operating condition variation difference values ​​collected in the historical time period and the adjustment coefficient influence set according to empirical rules are selected as training samples. The model is trained based on the XGBoost framework with the least square error as the objective function, and finally the decomposition layer influence sequence and the penalty parameter influence sequence are obtained after training.

[0039] Then, by multiplying the influence of the adjustment coefficient with the initial adjustment coefficient in the variational mode decomposition, a class of adjustment coefficients is obtained. The class of adjustment coefficients includes a class of adjustment coefficients based on the number of decomposition layers and a class of adjustment coefficients based on the penalty parameter.

[0040] Finally, the initial adjustment coefficients used to obtain the decomposition parameters are replaced with a class of adjustment coefficients to complete the variational mode decomposition and output the corresponding modal components.

[0041] By quantifying the difference between the influence index of operating condition changes and the threshold of the degree of influence of operating condition changes, a quantitative correlation between operating condition changes and decomposition parameters in variational mode decomposition is established. By inputting the difference value of operating condition changes into the pre-trained decomposition layer number influence sequence and penalty parameter influence sequence, the corresponding adjustment coefficient influence is obtained, realizing the adaptive adjustment of the decomposition layer number and penalty parameter. This method can dynamically optimize the variational mode decomposition parameters according to the actual operating condition changes, making the mode component separation more accurate and stable, effectively avoiding the mode aliasing and decomposition distortion problems that occur under traditional fixed parameters, thereby improving the accuracy and robustness of signal decomposition and providing more reliable input data for subsequent fault feature extraction and diagnosis.

[0042] Step 4: Determine whether the signal-to-noise ratio (SNR) of the original vibration signal is lower than the minimum SNR value. If so, change the Type I adjustment of the adaptive adjustment of the decomposition parameters to Type II adjustment based on the SNR. Otherwise, make a predictive adjustment judgment based on the SNR and the influence index of the working condition change to determine whether to make predictive adjustment of the decomposition parameters for the next preset time window. The aforementioned minimum SNR value is read from the preset database. It is usually preset by the preset staff based on historical data and empirical rules, and stored in the preset database for later retrieval.

[0043] Furthermore, the specific details of the second category of adjustments are as follows:

[0044] The signal-to-noise ratio offset is obtained by performing a difference operation between the original vibration signal's signal-to-noise ratio and its minimum value, followed by data normalization.

[0045] The signal-to-noise ratio offset and the difference value of the operating condition change are weighted and fused to obtain the fusion influence of the decomposition parameters. Specifically, the signal-to-noise ratio offset and the difference value of the operating condition change are multiplied by the weight factors extracted from the preset database and then summed. The weight factors include the first weight factor corresponding to the signal-to-noise ratio offset and the second weight factor corresponding to the difference value of the operating condition change.

[0046] The influence of the decomposition parameter fusion is compared in the preset decomposition parameter mapping table and penalty parameter mapping table to obtain the corresponding second-order adjustment coefficients, namely the second-order adjustment coefficient of the decomposition layer number and the second-order adjustment coefficient of the penalty parameter.

[0047] It should be added that the decomposition parameter mapping table and the penalty parameter mapping table are used to characterize the correlation between the fusion influence of the decomposition parameters and the second-order moderating coefficient. The construction process is as follows: the fusion influence of the decomposition parameters collected in the historical time period, and the second-order moderating coefficient preset by professionals according to empirical rules, are input into the initial data table constructed based on the logistic regression algorithm; then, the cross-entropy loss function is used as the optimization criterion, and the model is trained using the scikit-learn framework, finally generating the decomposition parameter mapping table and the penalty parameter mapping table.

[0048] Replace the first-class adjustment coefficients used for decomposition parameters with second-class adjustment coefficients to complete variational mode decomposition and output the corresponding modal components.

[0049] By introducing a weighted fusion mechanism based on the difference between signal-to-noise ratio offset and operating condition variations, adaptive optimization of variational mode decomposition parameters under low signal-to-noise ratio and complex operating conditions is achieved. Simultaneously, the influence of the decomposition parameters obtained through fusion calculation is compared in a preset mapping table to quickly determine the second type of adjustment coefficients, thereby dynamically correcting the number of decomposition layers and penalty parameters. This method not only comprehensively reflects the combined impact of signal quality and operating condition variations on decomposition accuracy but also significantly improves the algorithm's robustness and responsiveness to noise interference, ensuring the accuracy and stability of mode component separation and providing a more reliable signal foundation for subsequent fault feature extraction and intelligent diagnosis.

[0050] Figure 2 shows a flowchart of the predictive adjustment process provided in this embodiment of the invention. The specific logic is as follows: After normalizing the signal-to-noise ratio (SNR) and the operating condition change impact index, a numerical relationship judgment is performed: If the SNR is greater than the operating condition change impact index, and the deviation between the two falls within a preset first prediction interval, the predictive index score is marked as an invalid predictive score; if the deviation between the two is less than the minimum value of the preset first prediction interval, the predictive index score is marked as a first predictive score; otherwise, the predictive index score is marked as a second predictive score. If the SNR is less than the operating condition change impact index, and the deviation between the two falls within a preset second prediction interval, the predictive index score is marked as an invalid predictive score; if the deviation between the two is less than the minimum value of the preset second prediction interval, the predictive index score is marked as an invalid predictive score. If the predictive index score is positive, it is marked as the third predictive score; otherwise, it is marked as the fourth predictive score. If the signal-to-noise ratio equals the operating condition change influence index, the predictive index score is directly marked as an invalid predictive score. If the obtained predictive index score is invalid, no predictive adjustment judgment is made on the decomposition parameters of the next preset time window; otherwise, predictive adjustment is made on the decomposition parameters of the next preset time window based on the corresponding predictive index score. Through the above process, not only is it beneficial to improve the stability and response efficiency of the variational mode decomposition algorithm in continuous time series, and significantly enhance the decomposition results in terms of temporal continuity, accuracy and robustness, but it also provides more intelligent and stable technical support for online condition monitoring and intelligent diagnosis of rolling bearings.

[0051] It should be noted that the predictive adjustment judgment based on the signal-to-noise ratio and the impact index of operating condition changes is used to determine whether to predictively adjust the decomposition parameters for the next preset time window. The specific steps are as follows:

[0052] After normalizing the signal-to-noise ratio and the operating condition change impact index respectively, a numerical relationship is determined to obtain a predictive index score for assessing whether predictive adjustments should be made. This mainly includes the following situations:

[0053] In the first case, if the signal-to-noise ratio is greater than the operating condition change influence index and the deviation between the two falls within the preset first prediction interval, then the predictive index score is marked as an invalid predictive score. If the deviation between the two is less than the minimum value of the preset first prediction interval, then the predictive index score is marked as a first predictive score. Otherwise, the predictive index score is marked as a second predictive score. The deviation between the two represents the ratio of the difference between the signal-to-noise ratio and the operating condition change influence index to the operating condition change influence index.

[0054] In the second case, if the signal-to-noise ratio is less than the operating condition change impact index, and the deviation between the two falls within the preset second prediction interval, then the predictive index score is marked as an invalid predictive score. If the deviation between the two is less than the minimum value of the preset second prediction interval, then the predictive index score is marked as a third predictive score; otherwise, the predictive index score is marked as a fourth predictive score. The deviation between the two represents the ratio of the difference between the operating condition change impact index and the signal-to-noise ratio to the signal-to-noise ratio.

[0055] In the third case, if the signal-to-noise ratio equals the operating condition change impact index, then the predictive index score is directly marked as an invalid predictive score.

[0056] If the obtained predictive index score is an invalid predictive score, then no predictive adjustment judgment will be made on the decomposition parameters of the next preset time window; otherwise, predictive adjustment will be made on the decomposition parameters of the next preset time window based on the corresponding predictive index score.

[0057] It should be added that the aforementioned preset first prediction interval and preset second prediction interval are both preset intervals set by the staff, generally based on historical data and empirical rules; the predictive index score is a positive integer, and the values ​​of the invalid predictive score, first predictive score, second predictive score, third predictive score and fourth predictive score are preset by the staff based on different degrees of influence of different situations on predictive analysis. For example, the invalid predictive score is generally set to the value 0.

[0058] The specific methods of predictive adjustment are as follows:

[0059] The obtained predictive index scores are matched with the set predictive adjustment dataset to obtain the corresponding predictive adjustment ratio.

[0060] It should be added that by inputting the predictive index scores into the predictive regulation dataset, the corresponding predictive regulation ratio can be obtained. This dataset is used to fit the mapping relationship between the predictive index scores and the predictive regulation ratio. It is constructed as follows: in the initial dataset built based on the linear regression algorithm, the predictive index scores collected in the historical time period and the predictive regulation ratio set according to the empirical rules are selected as training samples. The least squares criterion is used and the model is trained based on the statsmodels framework to finally obtain the trained predictive regulation dataset.

[0061] The predictive adjustment ratio is multiplied by the decomposition parameters to obtain the predictive decomposition parameters, which include the number of predictive decomposition layers and the predictive penalty parameters.

[0062] When performing variational mode decomposition in the next preset time window, the initial decomposition parameters are replaced with predictive decomposition parameters to perform variational mode decomposition and output the corresponding mode components.

[0063] By introducing a predictive adjustment judgment mechanism based on the influence index of signal-to-noise ratio (SNR) and operating condition changes, this method achieves forward-looking optimization and dynamic self-learning of variational mode decomposition (MODED) parameters. Simultaneously, by normalizing the SNR and operating condition influence indices and establishing a numerical relationship judgment model, it generates predictive index scores to measure future adjustment trends. This allows for accurate identification of parameter adjustment needs in the next time window under different combinations of SNR and operating condition changes. The mechanism categorizes different predictive score types based on deviation intervals, distinguishing between operating condition-dominated, noise-dominated, and stable states, intelligently determining whether predictive parameter adjustment is necessary, thus avoiding ineffective or excessive parameter updates. Furthermore, compared to traditional static adjustment methods based solely on historical experience, this scheme possesses adaptive and forward-looking characteristics, enabling early detection of changes in operating conditions or signal quality trends and early optimization of decomposition parameter configuration for the next cycle. This method effectively improves the stability and response efficiency of the variational mode decomposition algorithm in continuous time series, significantly enhancing the temporal continuity, accuracy, and robustness of the decomposition results, providing more intelligent and stable technical support for online condition monitoring and intelligent diagnosis of rolling bearings.

[0064] Step 5: After the modal decomposition is completed, the decomposed modal components are output, and the enhanced vibration signal is obtained by reconstructing the signal based on the modal components. The signal is then input into the deep neural network for fault diagnosis, and the reliability of the fault diagnosis results is evaluated to determine whether to output the operating condition data sampling feedback command for the next preset time window.

[0065] Commonly used deep neural networks for fault diagnosis of rolling bearings include, but are not limited to, convolutional neural networks (such as LeNet and AlexNe), Bi-LSTM (Bidirectional Long Short-Term Memory) and Stacked LSTM (Stacked Long Short-Term Memory).

[0066] Taking convolutional neural networks as an example, historical vibration signals are obtained from historical databases and divided into training and test sets according to a preset ratio (e.g., 7:3 or 8:2). The training set is then input into the basic convolutional neural network model for training. After the training set has been trained, the test set is input into the basic convolutional neural network model for training. After training is completed, the corresponding deep neural network model for predicting rolling bearing failures is obtained.

[0067] The specific implementation method for obtaining the enhanced vibration signal through signal reconstruction is as follows:

[0068] Each modal component is numbered and its energy is obtained. At the same time, the corresponding weight coefficients are read from a preset database. The weight coefficients represent the relative importance of the corresponding modal component in signal reconstruction and are generally read from the preset database.

[0069] The energy, weight coefficient, and adjustment parameter of each modal component are substituted into the energy weight calculation formula to obtain the corresponding energy weight. The adjustment parameter is used to control the sensitivity of the energy difference of each modal component to the weight distribution.

[0070] The specific expression for the energy weight calculation formula is as follows:

[0071] ;

[0072] in, Indicates the first Weighting coefficients of each modal component, Indicates the first The energy of each modal component Indicates the adjustment parameter. Indicates the summation index, Indicates the numbering of the modal components. .

[0073] Substitute the weighting coefficients of each modal component into the reconstructed signal formula, output the reconstructed signal, and denot it as the enhanced vibration signal.

[0074] The specific expression for the reconstructed signal formula is as follows:

[0075] ;

[0076] in, Indicates enhanced vibration signal, Indicates the first One modal component, Indicates the weighting coefficient. Represents a time variable.

[0077] By introducing an energy-weighted signal reconstruction mechanism, the optimized integration of variational mode decomposition results is achieved. Simultaneously, by quantifying the energy of each modal component and combining it with weight coefficients and adjustment parameters from a pre-set database, an energy-weight calculation model is constructed. This model can adaptively adjust the contribution ratio of each modal component to the reconstructed signal based on its importance and energy distribution. This method effectively suppresses the interference of low-energy, noisy components on signal reconstruction, highlighting the influence of the main characteristic modes, thus obtaining a vibration signal with a higher signal-to-noise ratio and more significant features. This enhanced signal provides clearer and more stable input data for subsequent deep learning fault diagnosis models, significantly improving diagnostic accuracy and reliability.

[0078] The reliability of the fault diagnosis results is evaluated to determine whether to output the operating condition data sampling feedback command for the next preset time window. The specific process is as follows:

[0079] Obtain reliability assessment metrics that reflect the reliability characteristics of fault diagnosis results. These metrics include prediction confidence, prediction variance, and output entropy.

[0080] Specifically, prediction confidence represents the maximum class probability of the network output after temperature scaling, reflecting the model's confidence in the current prediction; prediction variance is the variance of multiple predictions by Monte Carlo Dropout or the ensemble model, representing the model's uncertainty; and output entropy is the information entropy of the output class distribution. Higher entropy indicates fuzzy classification and low confidence.

[0081] The system checks whether the reliability assessment indicators and the reference data extracted from the preset database meet the reliability judgment conditions. If the reliability conditions are met, the system outputs the fault diagnosis result; otherwise, it outputs the operating condition data sampling feedback instruction for the next preset time window. The operating condition data sampling feedback instruction prompts the preset staff to increase the sampling frequency of the operating condition data according to the preset sampling control ratio.

[0082] The reference data includes the lowest confidence level, the maximum prediction variance, and the highest output entropy. The reliability criteria are that the prediction confidence level is greater than the lowest confidence level, the prediction variance is less than the maximum prediction variance, and the output entropy is less than the highest output entropy.

[0083] By introducing a multi-dimensional reliability assessment mechanism, dynamic reliability determination of fault diagnosis results is achieved. By comprehensively considering three indicators—prediction confidence, prediction variance, and output entropy—the stability and consistency of the diagnostic model are more fully reflected. The results are compared with reference data in a pre-set database to determine whether the diagnostic results meet reliability requirements. When the reliability of the results is insufficient, the system automatically generates a working condition data sampling feedback instruction, prompting staff to increase the sampling frequency to obtain higher-quality data for model correction and re-diagnosis. This method effectively avoids misjudgments caused by abnormal data or model uncertainty, improves the accuracy and robustness of diagnostic results, and enables intelligent closed-loop monitoring and adaptive optimization of equipment operating status.

[0084] In this embodiment, by synchronously acquiring the original vibration signal, signal-to-noise ratio, and operating condition change data of the bearing within a preset time window, intelligent adjustment of decomposition parameters is achieved according to dynamic changes in operating conditions, thereby effectively improving the robustness and accuracy of signal decomposition and fault identification. Specifically, the scheme first generates an operating condition change influence index based on the operating condition change data, which can quantitatively characterize the change range of the bearing operating environment. When the operating condition change influence index exceeds a preset threshold, the system automatically performs a type I adaptive adjustment on the number of decomposition layers and penalty parameters of variational mode decomposition, so that the decomposition results can respond to operating condition disturbances in a timely manner. If the operating condition change is small, the parameters are kept stable to avoid unnecessary fluctuations, thereby improving the stability and computational efficiency of the algorithm. Secondly, by monitoring the signal-to-noise ratio of the vibration signal in real time, when the signal-to-noise ratio is lower than the set minimum value, the algorithm triggers a type II adjustment mechanism, which combines the signal-to-noise ratio to strengthen and correct the decomposition parameters, effectively compensating for the impact of noise on decomposition accuracy, and ensuring that reliable mode separation and feature preservation can still be achieved in low signal-to-noise ratio environments.

[0085] Furthermore, by jointly analyzing the signal-to-noise ratio and the influence index of operating condition changes, the system achieves predictive adjustment of parameters for the next time window, enabling the decomposition parameters to have forward-looking and self-learning capabilities. After decomposition, the obtained modal components are reconstructed into enhanced vibration signals, which are then input into a deep neural network for fault diagnosis. The system also achieves closed-loop optimization by feeding back sampling instructions through result reliability assessment.

[0086] In summary, this method not only significantly improves the decomposition accuracy and stability of variational mode decomposition under complex working conditions and noise backgrounds, but also enhances the generalization and intelligence level of the rolling bearing fault diagnosis model, ensuring the accuracy of early fault detection in rolling bearings.

[0087] Example 2:

[0088] In addition to ensuring that the main influencing factors are adjusted for Category II, the following are also included:

[0089] The signal-to-noise ratio offset is compared with the difference value of the operating condition change: if the signal-to-noise ratio offset is greater than the difference value of the operating condition change, and the difference between the two is higher than the preset difference value, then the first preset control amplitude is obtained based on the difference between the two.

[0090] It should be added that the difference between the signal-to-noise ratio offset and the difference in operating conditions is called the first control difference. The first control difference is input into the first control projection sequence to obtain the corresponding first preset control amplitude. This data sequence is used to fit the mapping relationship between the first control difference and the first preset control amplitude. The construction method is as follows: in the initial data sequence constructed based on the gradient boosting regression algorithm, the first control difference collected in the historical time period and the first preset control amplitude set according to empirical rules are selected as training samples. The model is trained based on the XGBoost framework with the least square error as the objective function, and finally the trained first control projection sequence is obtained.

[0091] The weights of the first preset control amplitude and the signal-to-noise ratio offset are multiplied to obtain the corresponding decomposition parameter fusion influence, and then the corresponding second-class control coefficient is obtained accordingly.

[0092] If the difference in operating conditions is greater than the signal-to-noise ratio offset, and the difference between the two is higher than the preset difference, then the second preset control amplitude is obtained based on the mapping of the difference between the two.

[0093] Similarly, the difference between the operating condition variation value and the signal-to-noise ratio offset is recorded as the second control difference. The second control difference is input into the second control projection sequence to obtain the corresponding second preset control amplitude. This data sequence is used to fit the mapping relationship between the second control difference and the second preset control amplitude. The construction method is as follows: in the initial data sequence constructed based on the gradient boosting regression algorithm, the second control difference collected in the historical time period and the second preset control amplitude set according to the empirical rules are selected as training samples. The model is trained based on the XGBoost framework with the least square error as the objective function, and finally the trained second control projection sequence is obtained.

[0094] The weights of the weighted fusion calculation based on the difference between the second preset control amplitude and the operating condition change are multiplied to obtain the corresponding decomposed parameter fusion influence, and then the corresponding second-class control coefficient is obtained accordingly.

[0095] In this embodiment, by introducing a comparison and dynamic adjustment mechanism between the signal-to-noise ratio offset and the difference in operating conditions, the second type of adaptive adjustment strategy for variational mode decomposition parameters is further improved. When there is a significant difference between the signal-to-noise ratio offset and the difference in operating conditions, the system obtains a corresponding preset adjustment amplitude based on the difference mapping, and dynamically determines the fusion influence of the decomposition parameters by combining the weights of the weighted fusion calculation, thereby obtaining the corresponding second type of adjustment coefficient and achieving precise optimization of the number of decomposition layers and penalty parameters. This method can automatically adjust the adjustment direction and amplitude according to the dominant influence of different operating conditions and noise levels, making the parameter adjustment both sensitive and stable. In addition, compared with the traditional static or single-factor driven adjustment method, this scheme achieves a two-way dynamic balance between signal quality and operating condition changes. Under low signal-to-noise ratio or drastic operating condition changes, it can still maintain clear mode component separation and reasonable energy concentration distribution. Through this mechanism, the robustness, adaptability, and decomposition accuracy of variational mode decomposition under complex operating conditions are significantly improved, providing a more reliable high-quality signal foundation for subsequent vibration signal reconstruction and fault feature identification.

[0096] According to embodiments of this disclosure, this disclosure also provides a rolling bearing fault diagnosis system based on a deep neural network. Figure 3 shows a schematic diagram of the structure of the rolling bearing fault diagnosis system based on a deep neural network provided in an embodiment of the present invention; including:

[0097] Data sampling module: Within the current preset time window, acquire the original vibration signal of the rolling bearing, its signal-to-noise ratio, and operating condition change data. The operating condition change data reflects the changes of the rolling bearing as the actual working environment changes.

[0098] Data processing and decomposition module: preprocesses the original vibration signal to obtain the corresponding vibration signal to be analyzed, and performs modal decomposition on the vibration signal to be analyzed based on the decomposition parameters in variational mode decomposition, including the number of decomposition layers and penalty parameters.

[0099] Operating condition influence decomposition module: During the variational mode decomposition process, an operating condition change influence index is generated based on the operating condition change data. It is then determined whether the operating condition change influence index is higher than the preset operating condition influence level threshold. If so, the decomposition parameters are adaptively adjusted based on the difference between the operating condition change influence index and the operating condition influence level threshold. Otherwise, variational mode decomposition continues.

[0100] Signal-to-noise ratio (SNR) adjustment module: Determines whether the SNR of the original vibration signal is lower than the minimum SNR value. If so, it changes the first-class adjustment of the adaptive adjustment of the decomposition parameters to a second-class adjustment based on the SNR. Otherwise, it makes a predictive adjustment judgment based on the SNR and the influence index of the working condition change to determine whether to make a predictive adjustment of the decomposition parameters for the next preset time window.

[0101] Fault diagnosis result output module: After the modal decomposition is completed, the decomposed modal components are output, and the enhanced vibration signal is obtained by reconstructing the signal based on the modal components. The signal is then input into the deep neural network for fault diagnosis, and the reliability of the fault diagnosis result is evaluated to determine whether to output the working condition data sampling feedback instruction for the next preset time window.

[0102] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0103] The above description is only an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A rolling bearing fault diagnosis method based on deep neural networks, characterized in that, The method includes: acquiring the original vibration signal of the rolling bearing, its signal-to-noise ratio, and operating condition change data within a current preset time window, wherein the operating condition change data reflects the changes in the rolling bearing as the actual working environment changes; preprocessing the original vibration signal to obtain the corresponding vibration signal to be analyzed, and performing mode decomposition on the vibration signal to be analyzed based on the decomposition parameters in variational mode decomposition, wherein the decomposition parameters include the number of decomposition levels and a penalty parameter; during the variational mode decomposition process, generating an operating condition change influence index based on the operating condition change data, and determining whether the operating condition change influence index is higher than a preset operating condition influence degree threshold; if so, adjusting the operating condition change influence index based on the difference between the operating condition change influence index and the operating condition influence degree threshold. The adaptive adjustment of the decomposition parameters is performed as a type I adjustment; otherwise, variational mode decomposition continues. It is determined whether the signal-to-noise ratio (SNR) of the original vibration signal is lower than the minimum SNR. If so, the type I adjustment of the adaptive adjustment of the decomposition parameters is changed to a type II adjustment based on the SNR. Otherwise, a predictive adjustment is performed based on the SNR and the influence index of operating condition changes to determine whether to perform predictive adjustment of the decomposition parameters for the next preset time window. After mode decomposition is completed, the decomposed modal components are output, and the enhanced vibration signal is reconstructed based on the modal components. This signal is then input into a deep neural network for fault diagnosis, and the reliability of the fault diagnosis results is evaluated to determine whether to output the operating condition data sampling feedback command for the next preset time window.

2. The rolling bearing fault diagnosis method based on deep neural networks as described in claim 1, characterized in that, The specific method for obtaining the operating condition change influence index is as follows: After normalizing the operating condition change data, standard operating condition change data is obtained. The operating condition change data includes speed, load torque, motor current, motor voltage, and bearing load. The standard operating condition change data is smoothed to obtain the operating condition change influence index. The operating condition change difference value is used to quantify the degree of influence of the actual operating condition of the rolling bearing on the variational mode decomposition.

3. The rolling bearing fault diagnosis method based on deep neural networks as described in claim 1, characterized in that, The adaptive adjustment of decomposition parameters based on the difference between the operating condition change influence index and the operating condition influence degree threshold is performed as follows: The difference between the operating condition change influence index and the operating condition influence degree threshold is quantified to obtain the operating condition change difference value; the operating condition change difference value is input into the pre-trained decomposition layer number influence sequence and the penalty parameter influence sequence, respectively, and the corresponding adjustment coefficient influence amounts, namely the decomposition layer number influence amount and the penalty parameter influence amount, are output respectively; the adjustment coefficient influence amounts are multiplied with the initial adjustment coefficients in variational mode decomposition to obtain a class of adjustment coefficients, which includes a class of adjustment coefficients for the decomposition layer number and a class of adjustment coefficients for the penalty parameter; the initial adjustment coefficients used to obtain the decomposition parameters are replaced with the class of adjustment coefficients to complete the variational mode decomposition and output the corresponding modal components.

4. The rolling bearing fault diagnosis method based on deep neural networks as described in claim 1, characterized in that, The specific content of the second type of adjustment is as follows: After performing a difference calculation between the signal-to-noise ratio of the original vibration signal and the minimum signal-to-noise ratio, the data is normalized to obtain the signal-to-noise ratio offset; the signal-to-noise ratio offset and the difference value of the working condition change are weighted and fused to obtain the fusion influence of the decomposed parameters. The influence of the decomposition parameter fusion is compared in the preset decomposition parameter mapping table and penalty parameter mapping table to obtain the corresponding second-type adjustment coefficients, namely the second-type adjustment coefficient of the decomposition layer number and the second-type adjustment coefficient of the penalty parameter. Replace the first-class adjustment coefficients used for decomposition parameters with second-class adjustment coefficients to complete variational mode decomposition and output the corresponding modal components.

5. The rolling bearing fault diagnosis method based on deep neural networks as described in claim 4, characterized in that, In addition to the two types of adjustments, the following steps are also included: comparing the signal-to-noise ratio (SNR) offset with the difference in operating conditions: if the SNR offset is greater than the difference in operating conditions, and the difference between the two is higher than a preset difference, then a first preset control amplitude is obtained based on the difference between the two; the first preset control amplitude and the SNR offset are weighted and fused together to calculate a product, thereby obtaining the corresponding decomposition parameter fusion influence, and thus obtaining the corresponding type II adjustment coefficient; if the difference in operating conditions is greater than the SNR offset, and the difference between the two is higher than a preset difference, then a second preset control amplitude is obtained based on the difference between the two; the second preset control amplitude and the difference in operating conditions are weighted and fused together to calculate a product, thereby obtaining the corresponding decomposition parameter fusion influence, and thus obtaining the corresponding type II adjustment coefficient.

6. The rolling bearing fault diagnosis method based on deep neural networks as described in claim 1, characterized in that, The method for determining whether to perform predictive adjustment of the decomposition parameters for the next preset time window based on signal-to-noise ratio (SNR) and the operating condition change influence index involves the following steps: After normalizing the SNR and the operating condition change influence index, a numerical relationship is determined to obtain a predictive index score for assessing whether to perform predictive adjustment. If the SNR is greater than the operating condition change influence index, and the deviation between the two falls within a preset first prediction interval, the predictive index score is marked as an invalid predictive score. If the deviation is less than the minimum value of the preset first prediction interval, the predictive index score is marked as a first predictive score; otherwise, the predictive index score is marked as a second predictive score. If the signal-to-noise ratio (SNR) is less than the operating condition change influence index, and the deviation between the two falls within the preset second prediction interval, then the predictive index score is marked as an invalid predictive score. If the deviation between the two is less than the minimum value of the preset second prediction interval, then the predictive index score is marked as a third predictive score; otherwise, the predictive index score is marked as a fourth predictive score. If the SNR is equal to the operating condition change influence index, then the predictive index score is directly marked as an invalid predictive score. If the obtained predictive index score is an invalid predictive score, then no predictive adjustment judgment is made on the decomposition parameters of the next preset time window; otherwise, predictive adjustment is made on the decomposition parameters of the next preset time window based on the corresponding predictive index score.

7. The rolling bearing fault diagnosis method based on deep neural networks as described in claim 6, characterized in that, The specific method of predictive adjustment is as follows: the obtained predictive index score is matched in the set predictive adjustment dataset to obtain the corresponding predictive adjustment ratio; the predictive adjustment ratio is multiplied with the decomposition parameters to obtain the predictive decomposition parameters, which include the number of predictive decomposition layers and the predictive penalty parameter. When performing variational mode decomposition in the next preset time window, the initial decomposition parameters are replaced with predictive decomposition parameters to perform variational mode decomposition and output the corresponding mode components.

8. The rolling bearing fault diagnosis method based on deep neural networks as described in claim 1, characterized in that, The specific implementation method for obtaining the enhanced vibration signal by reconstructing the signal is as follows: number each modal component and obtain the energy of each modal component, and at the same time read the corresponding weight coefficient from the preset database. The weight coefficient represents the relative importance of the corresponding modal component in the signal reconstruction. The energy, weight coefficient, and adjustment parameter of each modal component are substituted into the energy weight calculation formula to obtain the corresponding energy weight. The adjustment parameter is used to control the sensitivity of the energy difference of each modal component to the weight distribution. Substitute the weighting coefficients of each modal component into the reconstructed signal formula, output the reconstructed signal, and denot it as the enhanced vibration signal.

9. The rolling bearing fault diagnosis method based on deep neural networks as described in claim 1, characterized in that, The process of performing a reliability assessment on the fault diagnosis results to determine whether to output a working condition data sampling feedback instruction for the next preset time window is as follows: A reliability assessment index reflecting the reliability characteristics of the fault diagnosis results is obtained. The reliability assessment index includes prediction confidence, prediction variance, and output entropy. The reliability assessment index is compared with reference data extracted from a preset database to determine if they meet reliability criteria. If they do, the fault diagnosis result is output; otherwise, a working condition data sampling feedback instruction for the next preset time window is output. This working condition data sampling feedback instruction prompts preset personnel to increase the sampling frequency of the working condition data according to a preset sampling control ratio.

10. A rolling bearing fault diagnosis system based on deep neural networks, characterized in that, include: Data sampling module: Within the current preset time window, acquires the original vibration signal of the rolling bearing, its signal-to-noise ratio, and operating condition change data, wherein the operating condition change data reflects the changes of the rolling bearing with the actual working environment; Data processing and decomposition module: Preprocesses the original vibration signal to obtain the corresponding vibration signal to be analyzed, and performs modal decomposition on the vibration signal to be analyzed based on the decomposition parameters in variational mode decomposition, wherein the decomposition parameters include the number of decomposition layers and penalty parameters; Operating Condition Influence Decomposition Module: During variational mode decomposition, an operating condition change influence index is generated based on the operating condition change data. It is then determined whether the operating condition change influence index is higher than a preset operating condition influence threshold. If so, a type I adjustment is made to the adaptive adjustment of the decomposition parameters based on the difference between the operating condition change influence index and the operating condition influence threshold; otherwise, variational mode decomposition continues. Signal-to-Noise Ratio (SNR) Adjustment Module: It determines whether the SNR of the original vibration signal is lower than the minimum SNR value. If so, the type I adjustment to the adaptive adjustment of the decomposition parameters is changed to a type II adjustment based on the SNR; otherwise, a predictive adjustment judgment is made based on the SNR and the operating condition change influence index to determine whether to make predictive adjustments to the decomposition parameters for the next preset time window. Fault Diagnosis Result Output Module: After mode decomposition is completed, the decomposed modal components are output, and an enhanced vibration signal is obtained by reconstructing the signal based on the modal components. This signal is then input into a deep neural network for fault diagnosis, and the reliability of the fault diagnosis result is evaluated to determine whether to output a sampling feedback instruction for the operating condition data for the next preset time window.

Images