A system-oriented bearing fault prediction method using noise reduction autoencoder

Abstract

The present invention relates to a system-oriented bearing fault prediction method using a noise reduction autoencoder, which belongs to the technical field of bearing fault prediction and solves the problem in the prior art of failing to consider the influence of complex system noise on bearing fault prediction. The method comprises the following steps: obtaining parameters and operating condition information of a transmission chain to be tested, and determining a dangerous bearing as a target bearing; performing a dynamic simulation of the transmission chain to be tested, and obtaining simulation result data as simulated noise; performing a component-level degradation test, and obtaining a vibration signal during the degradation process as test data; performing signal mixing on the test data and the simulation result data, and obtaining a noisy signal; constructing and training a noise reduction autoencoder model, and outputting denoised bearing vibration data; constructing a fault prediction model, dividing the degradation process into a healthy stage, a slow degradation stage, and a fast degradation stage, and inputting the denoised bearing vibration data into the three independent models respectively to obtain the prediction value of the highest-precision model among the three independent models.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault prediction, and in particular to a system-oriented bearing fault prediction method using a noise reduction autoencoder. Background Art

[0002] Bearings are crucial components in all types of rotating machinery, providing support, positioning, and friction reduction. They withstand a wide range of cyclical loads, and their health plays a crucial role in the stable operation of the system. In practical applications, bearings are also a common and vulnerable part, attracting significant attention. Failure to promptly and effectively identify faults can pose significant economic and personnel risks. Therefore, to ensure the proper operation and safety of equipment and reduce operating and maintenance costs, rolling bearing fault prediction is of both theoretical and practical significance.

[0003] During operation, bearings generate corresponding vibration signals depending on their fault conditions. Their frequency domain characteristics and energy levels can serve as a basis for fault identification and prediction. Therefore, acquiring these vibration signals during testing and actual operation, combined with fault prediction models, can effectively control the extent of bearing damage and risk.

[0004] With the development of computers and artificial intelligence, fault prediction and analysis can be performed on industrial machinery through industrial information and data processing. Bearing health monitoring and prediction are gradually shifting from theoretical models to data-driven models, and preventive maintenance is also gradually transitioning to predictive maintenance. The development of intelligent operations and maintenance can effectively improve equipment operational stability and reliability.

[0005] Existing technology divides the entire bearing lifecycle into stages to distinguish fault levels. This is typically done in two or multiple stages. Both approaches require determining the health stage. The multi-stage classification is based on the severity of the bearing fault, as reflected in the statistical characteristics of the vibration signal or health indicators. A general fault prediction method determines the corresponding threshold and probability density based on the actual equipment requirements. When the confidence level exceeds the threshold, a fault is considered to have occurred or a more serious fault is imminent.

[0006] A Chinese patent application, publication number CN112632466A, published on April 9, 2021, titled "A Bearing Fault Prediction Method Based on Principal Component Analysis and Deep Bidirectional Long Short-Term Memory Network," discloses a method based on a deep bidirectional long short-term memory network. This network model processes input data that has been normalized and measured. However, this method only considers the vibration signal characteristics of the bearing itself and does not account for the complex system noise encountered in practical applications.

[0007] Therefore, an improved technical solution is needed to achieve the task of bearing fault prediction in complex systems taking into account the complex characteristics of noise. Summary of the Invention

[0008] In order to make up for the characteristics of existing fault prediction technology that only focuses on the bearing itself without considering system applications and has insufficient coverage of unhealthy conditions, one embodiment of the present invention takes into account the complex characteristics of the system's transmission chain noise and provides a system-oriented bearing fault prediction method using a denoising autoencoder. This method is based on a denoising autoencoder with a gated recurrent unit and an attention mechanism. It extracts the target bearing health indicator from the noisy signal health indicator and proposes a fault prediction model, and then realizes the degradation stage discrimination and degradation degree quantification covering the entire life cycle, so as to realize system-oriented fault prediction based on component-level degradation test data.

[0009] To achieve the above objectives, according to an embodiment of the present invention, a system-oriented bearing fault prediction method using a noise reduction autoencoder is provided, comprising the following steps:

[0010] Step S1, obtaining parameters and operating condition information of a transmission chain to be tested, and performing bearing life verification on bearings in the transmission chain to be tested, and determining dangerous bearings as target bearings, wherein the transmission chain to be tested is a system including bearings;

[0011] Step S2, performing dynamic simulation on the transmission chain to be tested, and obtaining simulation result data as simulated noise;

[0012] Step S3, performing a component-level degradation test on the target bearing, and obtaining a vibration signal during the degradation process as test data;

[0013] Step S4: performing signal mixing on the obtained test data and simulation result data to obtain a noisy signal, and establishing a bearing degradation process health index of the target bearing based on the test data and a bearing degradation process health index based on the noisy signal;

[0014] Step S5: constructing a denoising autoencoder model, using the bearing degradation process health index based on the test data and the bearing degradation process health index based on the noisy signal to train the denoising autoencoder model, making full use of the test data to accurately extract the target prediction quantity from the health index of the noisy signal, obtaining the target bearing health index extracted through denoising, and outputting it as denoised bearing vibration data;

[0015] Step S6: construct a fault prediction model, use the envelope difference method to divide the health status of the bearing, and divide the degradation process into a healthy stage, a slow degradation stage, and a rapid degradation stage. Three independent models are trained for data in different stages. The three independent models constitute a fault prediction model. The denoised bearing vibration data is input into the three independent models to generate three sets of prediction values ​​and evaluation indicators. The difference in evaluation indicators is used to determine the current degradation stage and output the prediction value of the highest-precision model among the three independent models to achieve fault prediction.

[0016] Optionally, the step S1 specifically includes:

[0017] Step S1.1, obtaining gear parameters, bearing characteristic parameters and design operating condition information of the transmission chain to be tested;

[0018] Step S1.2, perform gear stress calibration, calculate the meshing force according to the design working condition information and the gear parameters of the transmission chain, and then obtain the bearing radial load as the equivalent dynamic load, and use this as a basis to perform bearing life calibration, and obtain the bearing life of each bearing as the calibration result, and determine the temperature coefficient g by looking up the table T , hardness coefficient g H , reliability life correction factor a1, life correction factor a2 and operating condition factor a3, the bearing life is

[0019]

[0020] Where n is the operating speed, P is the equivalent dynamic load, C is the basic rated dynamic load, and ε is the life index;

[0021] Step S1.3, determine the dangerous bearings based on comprehensive comparison of the verification results.

[0022] Optionally, step S2 specifically includes:

[0023] Step S2.1, using the lumped parameter method to integrate the shaft mass of the transmission shaft in the transmission chain to be tested into the gear, and obtain the gear force element including the shaft mass;

[0024] Step S2.2, the system dynamics equation of the transmission chain to be tested is:

[0025]

[0026] Where M is the total mass of the system, C is the damping of the system, K is the stiffness matrix of the system, δ is the displacement of the system, and f is the external force vector of the system. The external force vector of the system is the input calculated according to the working condition, the stiffness matrix of the system is calculated according to the setting, and the displacement of the system is the result calculated according to the setting.

[0027] The damping of the system is the time-varying mesh damping, expressed as:

[0028]

[0029] Among them, ε g is the gear meshing damping ratio of the system, k(t) represents the time-varying meshing stiffness of the gears of the system, t is the time instant, I represents the moment of inertia of the gears of the system and the subscripts 1 and 2 represent different gears respectively, r b represents the gear pitch radius of the system;

[0030] In step S2.3, a system dynamics model is established in the simulation software, taking into account the time-varying meshing damping and stiffness in S2.2, completing the system dynamics calculation and extracting the vibration response of the target bearing position as the simulation result data, that is, as the simulated noise.

[0031] Optionally, step S3 specifically includes:

[0032] Step S3.1, selecting a representative working condition of the transmission chain to be tested from the obtained design working condition information and converting it into a bearing test load;

[0033] Step S3.2, calculating the single-point fault characteristic frequency of the target bearing in the transmission chain to be tested as a basis for selecting a vibration sensor;

[0034] Step S3.3, select a healthy bearing with the same model as the target bearing in the transmission chain to be tested as a test bearing, and select a constant test condition from the design operating condition information obtained in step S1 for the test bearing to operate until failure; in this process, the vibration signal of the test bearing is collected by the sensor according to a fixed single sampling time and sampling interval, and the data of the vibration signal is merged to obtain a vibration signal data set for the entire life cycle as test data.

[0035] Optionally, step S4 specifically includes:

[0036] Step S4.1, using cubic spline interpolation to match the sampling rates of the simulation result data and the test data to obtain a corrected noise, and copying the corrected noise signal to the same length as the test data to obtain a noise signal of equal length;

[0037] Step S4.2, mixing the equal-length noise signal obtained in the above steps and the test vibration signal in the time domain to obtain a noisy signal;

[0038] Step S4.3: extract health indicators from the test data and the noisy signal respectively, and evaluate the monotonicity and robustness of the extracted health indicators to obtain the health indicators with the best monotonicity and robustness;

[0039] In step S4.4, the obtained health index with the best monotonicity and robustness is used as an indirect prediction quantity, and the health index is calculated for the noisy signal and the test data respectively to obtain the health index of the bearing degradation process based on the noisy signal and the health index of the bearing degradation process based on the test data.

[0040] Optionally, the step S4.3 specifically includes:

[0041] The health index extraction includes selecting the root mean square value, kurtosis, peak value and skewness in the test data and noisy signal as candidate health indicators;

[0042] The monotonicity and robustness of the candidate health indicators are evaluated. For the dynamic sliding window, the input sequence X=(x1,x2,...,x i ,…,x N ), the root mean square value is:

[0043]

[0044] Kurtosis is expressed as:

[0045]

[0046] The peak value is expressed as:

[0047] Peak Value = max(X)

[0048] Skewness is expressed as:

[0049]

[0050] Where N is the sample size, μ is the mean of the distribution, σ is the standard deviation of the distribution, E is the expected value of the distribution, and x is the distribution of the expected value. i is the input sample, i is the sample number;

[0051] Monotonicity is calculated using the Spearman correlation coefficient, which is expressed as:

[0052]

[0053] Where ρ represents the Spearman correlation coefficient, which is used as an indicator of monotonicity;

[0054] The robustness is calculated as the average difference between the upper and lower envelopes, expressed as:

[0055]

[0056] Among them, u env represents the upper envelope calculation, l env represents the lower envelope calculation, x(t) represents the sample point corresponding to the local maximum or minimum value, t represents the index of the local maximum or minimum value, and rob represents the robustness index;

[0057] The monotonicity and robustness of the candidate health indicators are evaluated by the above monotonicity and robustness expressions, and the health indicator with the best monotonicity and robustness is obtained.

[0058] Optionally, step S5 specifically includes:

[0059] Step S5.1, obtaining a complete set of bearing degradation process health indicators based on test data and a set of bearing degradation process health indicators based on noisy signals, dividing them into training, validation, and test sets, and performing tensor quantization to obtain an input sequence;

[0060] Step S5.2: construct the gated recurrent unit of the main structure of the denoising autoencoder model. For the input sequence X=(x1, x2,…, x N ), the gated recurrent unit is expressed as:

[0061] r i =σ(W r ·[h i-1 ,x i ]+b z )

[0062] z i =σ(W z ·[h i-1 ,x i ]+b r )

[0063]

[0064]

[0065] Among them, r i is the reset gate output, z i is the update gate output, is the candidate hidden state, h i is the final hidden state, x i is the input sample, i represents the current input sequence number, σ is the sigmoid activation function, ⊙ represents element-wise multiplication, W r , W z and W are the weight matrices of the reset gate, update gate and candidate hidden state, respectively. r ,b z and b are the reset gate, update gate and bias of candidate hidden state respectively;

[0066] Step S5.3, based on the gated recurrent unit, build the multi-head attention mechanism of the main structure of the denoising autoencoder model to obtain the main structure of the denoising autoencoder model. For a given query vector Q, a given key vector Key and a given value vector V, as well as the corresponding weight matrix The multi-head attention mechanism can be expressed as:

[0067]

[0068] MultiHed=Concat(head1,head2,…,head j …,head h )W O

[0069]

[0070] Among them, d k is the dimension of the key vector Key, W O is the weight matrix used to update the input, softmax represents the normalization calculation function, Concat represents the vector merging calculation function, h represents the total number of attention heads, head J represents the jth attention head, j represents the sequence number of the attention head, Attention represents the self-attention mechanism, and MultiHead represents the multi-head attention;

[0071] Step S5.4, adding a neuron parameter release process and a linear layer to the main structure of the denoising autoencoder model built in step S5.3, and outputting the model prediction value;

[0072] Step S5.5: Based on the obtained model prediction value, the denoising autoencoder model constructed in steps S5.2 to S5.4 is trained by gradient descent using the MSE loss function and the Adam optimization algorithm to converge the model parameters. During the process, the changes in the MSE loss function are monitored and the optimal model is output;

[0073] In step S5.6, the obtained health index of the bearing degradation process based on the noisy signal is divided according to the given model sequence length, and input into the denoising autoencoder model to obtain the target bearing health index extracted by denoising, and output as denoised bearing vibration data.

[0074] Optionally, step S6 specifically includes:

[0075] Step S6.1, based on the health indicators extracted from the test data of the test bearing, the envelope difference method is used to analyze the trend thereof to obtain the envelope difference result;

[0076] Step S6.2, dividing the degradation process of the test bearing obtained in step S3 into stages, performing boundary identification based on the obtained envelope difference result, and dividing the degradation process into a healthy stage, a slow degradation stage, and a fast degradation stage;

[0077] In step S6.3, the full life cycle test data of the test bearing obtained in step S3 is split into three data sets according to the three stages mentioned above, and three independent models are trained respectively, so that each independent model is fully overfitted on the corresponding data set. The three trained independent models are used as fault prediction models;

[0078] In step S6.4, the denoised bearing vibration data of the target bearing in the transmission chain to be tested is input into the three trained independent models, and the prediction values ​​and evaluation indicators are output respectively. The evaluation indicators output by the three independent models are compared to identify the model with the highest prediction accuracy, obtain the prediction value of the model with the highest accuracy, and then identify the health stage of the prediction value.

[0079] Optionally, the overall application method of the system-oriented bearing fault prediction is: obtaining bearing component-level degradation test data and simulated system noise data, converting them into health indicators and dividing the degradation stages, and then training the denoising autoencoder and fault prediction model. For a set of real inputs in the application process, the denoising autoencoder first extracts the target bearing health indicators as the input of the fault prediction model, and then the fault prediction model uses this data to realize qualitative identification of the degradation stage and quantitative identification of the degradation degree, thereby realizing system-oriented transmission chain bearing fault prediction based on component degradation tests.

[0080] Compared with the prior art, an embodiment of the present invention provides a system-oriented bearing fault prediction method using a noise reduction autoencoder, which has at least the following advantages:

[0081] 1. Dynamic simulation was used to simulate system noise in actual transmission chain applications. This effectively avoided the high costs, long test cycles, complex and demanding test setups of system-level degradation testing, as well as the unpredictability of complex system randomness and the resulting multi-component degradation. This effectively controlled the difficulty of signal processing and improved the feasibility of target component processing. Simultaneously, the application of component-level degradation test data maximized the controllability of target signal variables and a high signal-to-noise ratio, minimizing the cost, difficulty, cycle time, and risk of testing, enabling the development of targeted and informed noise suppression algorithms.

[0082] 2. The present invention achieves the suppression of complex noise in noisy signals through a denoising autoencoder containing a gated recurrent unit and an attention mechanism. The processing effect of such noise with a frequency band that highly overlaps with the target signal, high energy density and non-stationary is significantly better than that of traditional filtering and wavelet threshold denoising methods, and can extract the health indicators of the target bearing from noisy signals with extremely low signal-to-noise ratio.

[0083] 3. The fault prediction model in the present invention includes the ability to process vibration signals of the bearing throughout its entire life cycle. Compared with the fault prediction method in the existing invention that uses statistical methods to identify whether the bearing deviates from a healthy state, it can further determine the current degradation stage and degree of degradation of the bearing, achieving more complete coverage of the degradation process, and therefore has a wider range of application scenarios.

[0084] The present invention's system-wide fault prediction approach, using a noise-reducing autoencoder, offers greater practical value than existing methods that only consider bearing fault prediction. In specific application scenarios, equipping each bearing with a high-precision sensor mounted close together is clearly not feasible, and actual drivetrain noise remains significant in such scenarios. Furthermore, the simulated noise generated through dynamic simulation is more representative of actual system applications than the white noise used in existing methods, providing more realistic noise characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. By referring to the drawings, the features and advantages of the present invention can be more clearly understood. The drawings are schematic and should not be understood as limiting the present invention in any way. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0086] Figure 1 This is a flowchart of a system-oriented bearing fault prediction method using a noise reduction autoencoder according to an embodiment of the present invention.

[0087] Figure 2 A schematic diagram of an implementation of a system-oriented bearing fault prediction method using a noise reduction autoencoder provided according to an embodiment of the present invention.

[0088] Figure 3 A schematic diagram of a process for obtaining component-level test data in an example of a system-oriented bearing fault prediction method using a denoising autoencoder provided in an embodiment of the present invention.

[0089] Figure 4 A schematic diagram of a process of obtaining simulated noise in an example of a system-oriented bearing fault prediction method using a denoising autoencoder provided in an embodiment of the present invention.

[0090] Figure 5 A schematic diagram of constructing a denoising autoencoder model in an example of a system-oriented bearing fault prediction method using a denoising autoencoder provided in an embodiment of the present invention.

[0091] Figure 6A schematic diagram of constructing a fault diagnosis model in an example of a system-oriented bearing fault prediction method using a denoising autoencoder provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0092] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present invention and the features therein can be combined with each other without conflict.

[0093] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0094] A system-oriented bearing fault prediction method using a noise reduction autoencoder according to an embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0095] Explanation of symbols:

[0096] g T Temperature coefficient of bearings;

[0097] g H The hardness coefficient of the bearing;

[0098] a1 bearing reliability life correction factor;

[0099] a2 bearing life correction factor;

[0100] a3 bearing operating condition coefficient;

[0101] L bearing life;

[0102] nThe operating speed of the bearing;

[0103] P is the equivalent dynamic load of the bearing;

[0104] C basic dynamic load rating of bearings;

[0105] ε is the life index of the bearing;

[0106] M system overall quality;

[0107] C system damping;

[0108] K is the stiffness matrix of the system;

[0109] δ system displacement;

[0110] f is the external force vector of the system;

[0111] εg The gear mesh damping ratio of the system;

[0112] k(t) is the time-varying mesh stiffness of the gears in the system;

[0113] Time t;

[0114] I. The moment of inertia of the gears in the system;

[0115] r b The system's gear pitch radius;

[0116] N sample size;

[0117] μ distribution mean;

[0118] σ distribution standard deviation;

[0119] E distribution expectation;

[0120] x i Input sample;

[0121] i Sample serial number;

[0122] X gives the input sequence;

[0123] ρSpearman correlation coefficient;

[0124] Rob robustness index;

[0125] u env Upper envelope calculation;

[0126] l env Lower envelope calculation;

[0127] x(t) represents the sample point corresponding to the local maximum or minimum value;

[0128] tThe local maximum or minimum index;

[0129] r i Reset gate output;

[0130] z i Update gate output;

[0131] Candidate hidden states;

[0132] h i Final hidden state;

[0133] σsigmoid activation function (S-shaped growth curve activation function);

[0134] W r are the weight matrices of the reset gates respectively;

[0135] W zUpdate the gate weight matrix;

[0136] W is the weight matrix of the candidate hidden state;

[0137] b r Reset the gate bias;

[0138] b z Update gate bias;

[0139] b Bias of candidate hidden states;

[0140] Q is the given query vector;

[0141] Key gives the key vector;

[0142] V given value vector;

[0143] The weight matrix of the query vector;

[0144] Weight matrix for a given key vector;

[0145] weight matrix of value vector;

[0146] d k Indicates the dimension of the key vector Key;

[0147] W O The weight matrix used to update the input;

[0148] Softmax normalization calculation function;

[0149] Concat vector merging calculation function;

[0150] head h hth attention head;

[0151] h the total number of attention heads;

[0152] head j The jth attention head;

[0153] j Attention head number;

[0154] Attention self-attention mechanism;

[0155] MultiHead multi-head attention;

[0156] y i Actual data for predicting step positions;

[0157] Model output value for the predicted step location.

[0158] refer to Figures 1-6 According to an embodiment of the present invention, a system-oriented bearing fault prediction method using a noise reduction autoencoder is provided, which takes into account the complex characteristics of the noise of the bearing transmission chain and the system and its impact on the fault, and includes the following steps.

[0159] Step S1. Obtain parameters and operating condition information for a transmission chain under test, perform bearing life verification on bearings in the transmission chain under test, and identify at-risk bearings as target bearings. The transmission chain under test is a system including bearings. At-risk bearings are those most prone to failure within the entire transmission chain under test.

[0160] comparison Figure 1 In this embodiment, the bearing life verification method in the transmission chain to be tested includes the following steps.

[0161] Step S1.1: Obtain the gear parameters, bearing characteristic parameters, and design operating condition information of the transmission chain to be tested. Gear parameters for the transmission chain may include module, number of teeth, and tooth width; bearing characteristic parameters may include inner and outer diameters, number of rods, and pitch radius; and design operating condition information may include torque and speed. These parameters can be obtained by presetting. Gear parameters and bearing characteristic parameters are geometric and positional parameters and can be obtained from a designed 3D model of the transmission chain. Design operating condition information can be obtained from the transmission chain design.

[0162] Step S1.2, based on this, the gear stress is checked, the meshing force is calculated according to the design working condition information and the gear parameters of the transmission chain, and the bearing radial load is obtained as the equivalent dynamic load. Based on this, the bearing life check is carried out and the bearing life of each bearing is obtained as the check result. The bearing temperature coefficient g can be determined by looking up the table T , hardness coefficient g H , reliability life correction factor a1, life correction factor a2 and operating condition factor a3, the bearing life can be expressed as

[0163]

[0164] Where n is the bearing's operating speed, P is the bearing's equivalent dynamic load, C is the bearing's basic dynamic load rating, and ε is the bearing's life exponent. By adjusting the equivalent dynamic load, the expected test time can be controlled to verify bearing life.

[0165] In step S1.3, based on the comprehensive comparison of the verification results, the at-risk bearing is identified as the target bearing. Furthermore, the sensor installation position close to the at-risk bearing is determined based on the pre-set geometric features of the three-dimensional digital model of the transmission chain casing. This sensor installation position serves as the spatial location for acquiring system-simulated noise and the foundation for system-level fault prediction.

[0166] Step S2: Perform dynamic simulation on the transmission chain to be tested, and obtain simulation result data as simulated noise.

[0167] comparison Figure 4 , the transmission chain dynamics simulation method in this embodiment includes the following steps.

[0168] In step S2.1, the lumped parameter method is used to integrate the shaft mass of the transmission chain under test within the gears, generating a gear force element that includes the shaft mass. Three assumptions are made for multibody dynamics simulations: first, gear flexibility is ignored; second, only rotational and translational degrees of freedom along the contact path are considered; and third, since gear meshing is the primary source of noise, the support bearings are treated as linear spring-damper elements.

[0169] In step S2.2, the system dynamics equation of the transmission chain to be tested can be expressed as:

[0170]

[0171] Where M is the total mass of the system, C is the system damping, K is the system stiffness matrix, δ is the system displacement, and f is the system's external force vector. The system's external force vector is the input calculated based on the operating conditions, the system's stiffness matrix is ​​calculated based on the settings, and the system's displacement is the result calculated based on the settings.

[0172] The damping of the system can be time-varying meshing damping, which can be expressed as:

[0173]

[0174] Among them, ε g is the gear meshing damping ratio of the system, which ranges from 0.03 to 0.17, k(t) represents the time-varying meshing stiffness of the gears of the system, t is the time, I represents the moment of inertia of the gears of the system, and the subscripts 1 and 2 represent different gears, respectively, r b Represents the gear pitch radius of the system.

[0175] Step S2.3, establish a system dynamics model in the simulation software, consider the time-varying meshing damping and stiffness in S2.2, complete the system dynamics calculation and extract the vibration response of the target bearing position (i.e., simulation result data) as simulated noise.

[0176] Step S3: Perform a component-level degradation test on the target bearing determined in step S1, and obtain a vibration signal of the test bearing during the degradation process as test data. The test data is component-level test data.

[0177] comparison Figure 3 , the degradation test signal acquisition method in this embodiment includes the following steps.

[0178] Step S3.1: Select representative working conditions of the transmission chain to be tested from the obtained design working condition information and convert them into bearing test loads.

[0179] In step S3.2, calculate the characteristic frequency of single-point faults of the target bearing in the transmission chain under test as a basis for selecting a vibration sensor. Experience shows that faults encountered during degradation testing are often multiple faults, with characteristic peaks typically occurring between 2 and 4 times the single-point fault frequency. Therefore, according to the Nyquist sampling theorem, the sensor sampling frequency should be at least 8 times the single-point fault frequency. Other factors in selecting a vibration sensor include operating temperature and measuring range.

[0180] Step S3.3: According to the test conditions and the requirements of the above steps, a tester is set up. A healthy bearing of the same model as the target bearing in the transmission chain to be tested is selected as the test bearing. The test bearing is sequentially operated under constant test conditions selected from the design operating condition information obtained in step S1 until failure. During this process, the vibration signal of the test bearing is collected by a sensor according to a fixed single sampling duration and sampling interval, and the vibration signal data is combined to obtain a vibration signal data set during the degradation process (i.e., the entire life cycle) as the test data. The tester may include a drive motor, a loading system, and a sensor.

[0181] In step S4, the obtained test data and simulation result data are mixed to obtain a noisy signal. A bearing degradation process health indicator based on the test data and a bearing degradation process health indicator based on the noisy signal are then established for the target bearing. These established bearing degradation process health indicators based on the test data and the noisy signal serve as input to the denoising autoencoder model in step S5.

[0182] comparison Figure 2 ,The signal mixing and health index establishment method in this embodiment includes the following.

[0183] In step S4.1, cubic spline interpolation is used to match the sampling rates of the simulation results and test data to obtain corrected noise. This corrected noise signal is then replicated to the same length as the test data to obtain a noise signal of equal length. This noise signal is then used to achieve noise coverage over the entire bearing lifecycle. Due to computational limitations, simulated noise often differs from the measured test data in terms of sampling rate and signal length. The above processing method can address this difference, resulting in more accurate results.

[0184] Step S4.2: Mix the equal-length noise signal and the test vibration signal obtained in the above steps in the time domain to obtain a noisy signal.

[0185] In step S4.3, health indicators are extracted from the test data and the noisy signal, and the extracted health indicators are evaluated for monotonicity and robustness. This step can address issues such as excessive signal fluctuations and unclear degradation features in direct fault prediction.

[0186] The health index extraction includes selecting the root mean square value, kurtosis, peak value and skewness in the test data and noisy signals as candidate health indicators.

[0187] Next, the monotonicity and robustness of the candidate health indicators are evaluated. For the dynamic sliding window, the input sequence X=(x1,x2,...,x i , … ,x N ), the root mean square value can be expressed as:

[0188]

[0189] Kurtosis can be expressed as:

[0190]

[0191] The peak value can be expressed as:

[0192] Peak Value = max(X)

[0193] Skewness can be expressed as:

[0194]

[0195] Where N is the sample size, μ is the mean of the distribution, σ is the standard deviation of the distribution, E is the expected value of the distribution, and x is the distribution of the expected value. i is the input sample, and i is the sample number.

[0196] Monotonicity is calculated using the Spearman correlation coefficient, which can be expressed as:

[0197]

[0198] Here, ρ represents the Spearman correlation coefficient, which is used as an indicator of monotonicity.

[0199] The robustness is calculated as the average difference between the upper and lower envelopes, which can be expressed as:

[0200]

[0201] Among them, u env represents the upper envelope calculation, l env represents the lower envelope calculation, x(t) represents the sample point corresponding to the local maximum or minimum value, t is the index of the local maximum or minimum value, and rob represents the robustness index.

[0202] The monotonicity and robustness of the candidate health indicators are evaluated by the above monotonicity and robustness expressions, and the health indicator with the best monotonicity and robustness is obtained.

[0203] In step S4.4, the obtained health index with the best monotonicity and robustness is used as an indirect prediction quantity, and the health index is calculated for the noisy signal and the test data respectively to obtain the health index of the bearing degradation process based on the noisy signal and the health index of the bearing degradation process based on the test data.

[0204] Step S5: Construct a denoising autoencoder model. This model is trained using the bearing degradation process health indicator based on the test data and the bearing degradation process health indicator based on the noisy signal. The model fully utilizes the test data to accurately extract the target prediction quantity from the health indicator of the noisy signal, thereby outputting the target bearing health indicator after denoising. The output of this denoising autoencoder model serves as the input to the fault prediction model. Specifically, the health indicator of the noisy signal is restored through denoising to obtain the target bearing health indicator.

[0205] comparison Figure 5 , the method for constructing a denoising autoencoder model in this embodiment includes the following steps.

[0206] Step S5.1, obtain a complete set of bearing degradation process health indicators based on experimental data and bearing degradation process health indicators based on noisy signals, and divide them into training, verification and test sets, and perform tensor quantization to obtain the input sequence for adaptive model structure and graphics processing unit (GPU) accelerated training.

[0207] Step S5.2: Build the gated recurrent unit of the main structure of the denoising autoencoder model. … ,x N ), the gated recurrent unit can be expressed as:

[0208] r i =σ(W r ·[h i-1 ,x i ]+b z )

[0209] z i =σ(W z ·[h i-1 ,x i ]+b r )

[0210]

[0211] Among them, r i is the reset gate output, z i is the update gate output, is the candidate hidden state, h i is the final hidden state, x i is the input sample, i represents the current input sequence number, and N is the total number of samples. σ is the sigmoid activation function (S-type growth curve activation function), ⊙ represents element-wise multiplication, and W r , W z and W are the weight matrices of the reset gate, update gate and candidate hidden state, respectively. r ,b z and b are the reset gate, update gate and bias of the candidate hidden state respectively.

[0212] Step S5.3, based on the gated recurrent unit, build the multi-head attention mechanism of the main structure of the denoising autoencoder model to obtain the main structure of the denoising autoencoder model. For a given query vector Q, a given key vector Key and a given value vector V, and the corresponding weight matrix The multi-head attention mechanism can be expressed as:

[0213]

[0214] MultiHead=Concat(head1,head2,…,head j …,head h )W O

[0215]

[0216] Among them, d k is the dimension of the key vector Key, W O is the weight matrix used to update the input, ssoftmax represents the normalization calculation function, Concat represents the vector merging calculation function, h represents the total number of attention heads, head j represents the j-th attention head, j represents the sequence number of the attention head, Attention represents the self-attention mechanism, and MultiHead represents multi-head attention.

[0217] In step S5.4, a neuron parameter dropout process and a linear layer are added to the main structure of the denoising autoencoder model built in step S5.3, and the model prediction value is output. This step can improve the robustness of the denoising autoencoder model.

[0218] In step S5.5, based on the obtained model prediction value, the denoising autoencoder model built in steps S5.2 to S5.4 is trained using the MSE loss function (mean square error loss function) and the Adam optimization algorithm (stochastic optimization method with adaptive momentum) through gradient descent method to make the model parameters converge. During the process, the changes in the MSE loss function are monitored and the optimal model is output. The MSE loss function can be expressed as:

[0219]

[0220] Among them, y i is the true value, that is, the actual data of the predicted step position, which comes from the data set, and is the model prediction value, that is, the model output value of the prediction step position, i represents the i-th sample, and N represents the total number of samples.

[0221] In step S5.6, the bearing degradation process health indicator obtained based on the noisy signal is divided according to the given model sequence length and input into the denoising autoencoder model to obtain the target bearing health indicator extracted through denoising as output, which serves as denoised bearing vibration data. In subsequent steps, this denoised bearing vibration data can be used as input to the fault prediction model.

[0222] Step S6: construct a fault prediction model, use the envelope difference method to divide the health status of the bearing, divide the degradation process of the test bearing into a healthy stage, a slow degradation stage, and a rapid degradation stage, train three independent models for data in different stages, and these three independent models constitute a fault prediction model. Input the denoised bearing vibration data into the three trained independent models respectively to generate three sets of prediction values ​​and evaluation indicators. Use the difference in evaluation indicators to determine the current degradation stage, and output the prediction value of the highest-precision model among the three independent models to achieve fault prediction.

[0223] comparison Figure 6 , the fault prediction method of this embodiment includes the following steps.

[0224] Step S6.1: Based on the health indicators extracted from the test data of the test bearing, the envelope difference method is used to analyze the trend thereof to obtain the envelope difference result.

[0225] In step S6.2, the degradation process of the test bearing obtained in step S3 is divided into stages. Boundaries are identified based on the envelope difference results, and the stages are divided into healthy, slowly degraded, and rapidly degraded stages. The bearing undergoes a running-in process during the initial operation, which is reflected by fluctuations in the health indicator curve. The first non-zero value of the envelope difference after entering the normal operation phase is identified as the fault initiation point, also known as the first predicted time (FPT). The maximum value of the envelope difference during the identification process is used as the basis for reflecting sudden changes and further deterioration in the bearing health state. Using the identified fault initiation point and the maximum value of the envelope difference as two key points, the bearing health state is divided into three stages: healthy, slowly degraded, and rapidly degraded.

[0226] In step S6.3, the full life cycle test data of the test bearing obtained from step S3 is split into three data sets according to the above three stages, and three independent models are trained respectively, so that each independent model is fully overfitted on the corresponding data set. The three trained independent models are used as fault prediction models.

[0227] In step S6.4, the denoised vibration data of the target bearing in the transmission chain to be tested is input into the three trained independent models, and the prediction values ​​and evaluation indicators are output respectively. The prediction values ​​of the model with the highest accuracy among the three independent models are obtained by comparison. For a given input, the three independent models of the fault prediction model output prediction values ​​and evaluation indicators respectively. Among them, the given input can be, for example, the denoised bearing vibration data obtained in step S5. The evaluation indicator is the absolute percentage error, which can be expressed as follows: since the models are fully overfitted on the data of their respective health stages, only a single model shows a higher prediction accuracy for a single input. Therefore, by comparing the evaluation indicators output by the three independent models, the model with the highest prediction accuracy can be identified, and then the health stage of the corresponding data can be identified to achieve qualitative judgment of the degradation state; and the prediction value of the highest accuracy model is output, that is, quantitative analysis of the current health indicator is achieved.

[0228] Specifically, in order to ensure the differences in evaluation indicators between different models, the fault prediction model described in steps S6.3 and S6.4 should not have too strong generalization ability, which will bring difficulties to the discrimination of degradation stages. Therefore, a gated recurrent unit model is used and the neuron parameter release operation is not performed, and the predicted value is directly output through the linear layer.

[0229] According to an embodiment of the present invention, a system-oriented bearing fault prediction method using a denoising autoencoder is generally as follows: bearing component-level degradation test data and simulated system noise data are obtained, converted into health indicators and degradation stage divisions are performed, and then the denoising autoencoder and fault prediction model are trained. For a set of real inputs in the application process, the denoising autoencoder first extracts the test bearing health indicators as the input of the fault prediction model, and then the fault prediction model uses this data to realize qualitative identification of the degradation stage and quantitative identification of the degradation degree, thereby realizing system-oriented transmission chain bearing fault prediction based on component degradation tests.

[0230] An exemplary embodiment of a system-oriented bearing fault prediction method using a denoising autoencoder according to an embodiment of the present invention is described below.

[0231] This exemplary embodiment uses data ranging in length from 170,000 to 300,000, varying depending on the specific degradation time of the bearing. The data is divided into 60% training data, 20% validation data, and 20% test data. The denoising autoencoder uses a sliding window size of 1024 and a superior hidden unit count of 512, while the fault prediction model uses a superior hidden unit count of 128. Verified accuracy rates for both models are 95% and 97%, respectively, for an overall accuracy of 92%. Therefore, the method provided by this embodiment of the present invention can effectively predict faults while accounting for transmission system noise.

[0232] All of the above optional technical solutions can be combined in any way to form optional embodiments of the present application, and will not be described in detail here.

[0233] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0234] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A system-oriented bearing fault prediction method using a noise reduction autoencoder, characterized in that: include: Step S1, obtaining parameters and operating condition information of a transmission chain to be tested, and performing bearing life verification on bearings in the transmission chain to be tested, and determining dangerous bearings as target bearings, wherein the transmission chain to be tested is a system including bearings; Step S2, performing dynamic simulation on the transmission chain to be tested, and obtaining simulation result data as simulated noise; Step S3, performing a component-level degradation test on the target bearing, and obtaining a vibration signal during the degradation process as test data; Step S4: performing signal mixing on the obtained test data and simulation result data to obtain a noisy signal, and establishing a bearing degradation process health index of the target bearing based on the test data and a bearing degradation process health index based on the noisy signal; Step S5: constructing a denoising autoencoder model, using the bearing degradation process health index based on the test data and the bearing degradation process health index based on the noisy signal to train the denoising autoencoder model, making full use of the test data to accurately extract the target prediction quantity from the health index of the noisy signal, obtaining the target bearing health index extracted through denoising, and outputting it as denoised bearing vibration data; Step S6: Construct a fault prediction model. Use the envelope difference method to divide the bearing's health status, dividing the degradation process into a healthy stage, a slow degradation stage, and a rapid degradation stage. Train three independent models for the data in different stages. These three independent models constitute the fault prediction model. Input the de-noised bearing vibration data into the three independent models to generate three sets of prediction values ​​and evaluation indicators. Use the difference in the evaluation indicators to determine the current degradation stage and output the prediction value of the highest-precision model among the three independent models to achieve fault prediction. Wherein, step S2 includes: Step S2.1, using the lumped parameter method to integrate the shaft mass of the transmission shaft in the transmission chain to be tested into the gear, and obtain the gear force element including the shaft mass; Step S2.2: Establish the system dynamics equation of the transmission chain to be tested, and use the time-varying meshing damping as the system damping, which can be expressed as: Among them, ε g is the gear meshing damping ratio of the system, k(t) represents the time-varying meshing stiffness of the gears of the system, t is the time, I represents the moment of inertia of the gears of the system, r b Indicates the gear pitch radius of the system, with subscripts 1 and 2 representing different gears respectively; In step S2.3, a system dynamics model is established in the simulation software, the time-varying meshing damping and stiffness in step S2.2 are considered, the system dynamics calculation is completed, and the vibration response of the target bearing position is extracted as the simulation result data, that is, as the simulated noise.

2. The system-oriented bearing fault prediction method using a noise reduction autoencoder according to claim 1 is characterized in that: The step S1 specifically includes: Step S1.1, obtaining gear parameters, bearing characteristic parameters and design operating condition information of the transmission chain to be tested; Step S1.2, perform gear stress calibration, calculate the meshing force according to the design working condition information and the gear parameters of the transmission chain, and then obtain the bearing radial load as the equivalent dynamic load, and use this as a basis to perform bearing life calibration, and obtain the bearing life of each bearing as the calibration result, and determine the temperature coefficient g by looking up the table T , hardness coefficient g H , reliability life correction factor a1, life correction factor a2 and operating condition factor a3, the bearing life is Where n is the operating speed, P is the equivalent dynamic load, C is the basic rated dynamic load, and ε is the life index; Step S1.3, determine the dangerous bearings based on comprehensive comparison of the verification results.

3. The system-oriented bearing fault prediction method using a noise reduction autoencoder according to claim 2, characterized in that: The step S3 specifically includes: Step S3.1, selecting a representative working condition of the transmission chain to be tested from the obtained design working condition information and converting it into a bearing test load; Step S3.2, calculating the single-point fault characteristic frequency of the target bearing in the transmission chain to be tested as a basis for selecting a vibration sensor; Step S3.3, select a healthy bearing with the same model as the target bearing in the transmission chain to be tested as a test bearing, and select a constant test condition from the design operating condition information obtained in step S1 for the test bearing to operate until failure; in this process, the vibration signal of the test bearing is collected by the sensor according to a fixed single sampling time and sampling interval, and the data of the vibration signal is merged to obtain a vibration signal data set for the entire life cycle as test data.

4. The system-oriented bearing fault prediction method using a noise reduction autoencoder according to claim 3 is characterized in that: The step S4 specifically includes: Step S4.1, using cubic spline interpolation to match the sampling rates of the simulation result data and the test data to obtain a corrected noise, and copying the corrected noise signal to the same length as the test data to obtain a noise signal of equal length; Step S4.2, mixing the equal-length noise signal obtained in the above steps and the test vibration signal in the time domain to obtain a noisy signal; Step S4.3: extract health indicators from the test data and the noisy signal respectively, and evaluate the monotonicity and robustness of the extracted health indicators to obtain the health indicators with the best monotonicity and robustness; In step S4.4, the obtained health index with the best monotonicity and robustness is used as an indirect prediction quantity, and the health index is calculated for the noisy signal and the test data respectively to obtain the health index of the bearing degradation process based on the noisy signal and the health index of the bearing degradation process based on the test data.

5. The system-oriented bearing fault prediction method using a noise reduction autoencoder according to claim 4 is characterized in that: The step S4.3 specifically includes: The health index extraction includes selecting the root mean square value, kurtosis, peak value and skewness in the test data and noisy signal as candidate health indicators; The monotonicity and robustness of the candidate health indicators are evaluated. For the dynamic sliding window, the input sequence X=(x1,x2,...,x i ,…,x N ), the root mean square value is: Kurtosis is expressed as: The peak value is expressed as: Peak Value = max(X) Skewness is expressed as: Where N is the sample size, μ is the mean of the distribution, σ is the standard deviation of the distribution, E is the expected value of the distribution, and x is the distribution of the expected value. i is the input sample, i is the sample number; Monotonicity is calculated using the Spearman correlation coefficient, which is expressed as: Where ρ represents the Spearman correlation coefficient, which is used as an indicator of monotonicity; The robustness is calculated as the average difference between the upper and lower envelopes, expressed as: Among them, u env represents the upper envelope calculation, l env represents the lower envelope calculation, x(t) represents the sample point corresponding to the local maximum or minimum value, t represents the index of the local maximum or minimum value, and rob represents the robustness index; The monotonicity and robustness of the candidate health indicators are evaluated by the above monotonicity and robustness expressions, and the health indicator with the best monotonicity and robustness is obtained.

6. The system-oriented bearing fault prediction method using a noise reduction autoencoder according to claim 4, characterized in that: The step S5 specifically includes: Step S5.1, obtaining a complete set of bearing degradation process health indicators based on test data and a set of bearing degradation process health indicators based on noisy signals, dividing them into training, validation, and test sets, and performing tensor quantization to obtain an input sequence; Step S5.2: construct the gated recurrent unit of the main structure of the denoising autoencoder model. For the input sequence X=(x1, x2,…, x N ), the gated recurrent unit is expressed as: r i =σ(W r ·[h i-1 ,x i ]+b z ) z i =σ(W z ·[h i-1 ,x i ]+b r ) Among them, r i is the reset gate output, z i is the update gate output, is the candidate hidden state, h i is the final hidden state, x i is the input sample, i represents the current input sequence number, σ is the sigmoid activation function, ⊙ represents element-wise multiplication, W r , W z and W are the weight matrices of the reset gate, update gate and candidate hidden state, respectively. r ,b z and b are the reset gate, update gate and bias of candidate hidden state respectively; Step S5.3, based on the gated recurrent unit, build the multi-head attention mechanism of the main structure of the denoising autoencoder model to obtain the main structure of the denoising autoencoder model. For a given query vector Q, a given key vector Key and a given value vector V, as well as the corresponding weight matrix The multi-head attention mechanism can be expressed as: MultiHead=Concat(head1,Head2,…,head j …,head h )W O Among them, d k is the dimension of the key vector Key, W O is the weight matrix used to update the input, softmax represents the normalization calculation function, Concat represents the vector merging calculation function, h represents the total number of attention heads, head j represents the jth attention head, j represents the sequence number of the attention head, Attention represents the self-attention mechanism, and MultiHead represents the multi-head attention; Step S5.4, adding a neuron parameter release process and a linear layer to the main structure of the denoising autoencoder model built in step S5.3, and outputting the model prediction value; Step S5.5: Based on the obtained model prediction value, the denoising autoencoder model constructed in steps S5.2 to S5.4 is trained by gradient descent using the MSE loss function and the Adam optimization algorithm to converge the model parameters. During the process, the changes in the MSE loss function are monitored and the optimal model is output; In step S5.6, the obtained health index of the bearing degradation process based on the noisy signal is divided according to the given model sequence length, and input into the denoising autoencoder model to obtain the target bearing health index extracted by denoising, and output as denoised bearing vibration data.

7. The system-oriented bearing fault prediction method using a noise reduction autoencoder according to claim 6, characterized in that: The step S6 specifically includes: Step S6.1, based on the health indicators extracted from the test data of the test bearing, the envelope difference method is used to analyze the trend thereof to obtain the envelope difference result; Step S6.2, dividing the degradation process of the test bearing obtained in step S3 into stages, performing boundary identification based on the obtained envelope difference result, and dividing the degradation process into a healthy stage, a slow degradation stage, and a fast degradation stage; In step S6.3, the full life cycle test data of the test bearing obtained in step S3 is split into three data sets according to the three stages mentioned above, and three independent models are trained respectively, so that each independent model is fully overfitted on the corresponding data set. The three trained independent models are used as fault prediction models; In step S6.4, the denoised bearing vibration data of the target bearing in the transmission chain to be tested is input into the three trained independent models, and the prediction values ​​and evaluation indicators are output respectively. The evaluation indicators output by the three independent models are compared to identify the model with the highest prediction accuracy, obtain the prediction value of the model with the highest accuracy, and then identify the health stage of the prediction value.

Images