Method and system for wear monitoring of self-lubricating hinge bearing

Abstract

The application provides a kind of self-lubricating hinge bearing wear detection method and system, adopt convolution neural network (CNN) and the model of the combination of bidirectional gate cycle unit network (Bi-GRU). The application installs acoustic emission sensor and torque sensor in the key position of hinge bearing, and collects relevant signals in real time. The collected acoustic emission and torque data first go through a series of data processing steps, including data standardization, signal filtering, signal decomposition, feature extraction and feature selection steps, to ensure data quality and analyzability. These processed data are further converted into time-frequency features and input into the convolution neural network and long short-term memory network. The application significantly enhances the comprehensive analysis capability of data through the fusion of multi-sensor information, achieving high-precision identification of wear state. The model structure of the application optimizes the information integration process, ensuring the comprehensiveness and accuracy of wear detection, and improving the maintenance and operation efficiency of equipment.

Description

Technical Field

[0001] This invention relates to the field of fault condition monitoring technology, specifically to a method and system for monitoring the wear of self-lubricating hinge bearings, and more specifically to a method for monitoring the wear of self-lubricating hinge bearings based on multi-sensor fusion of acoustic emission and torque. Background Technology

[0002] Mechanical fault condition monitoring technology is used to understand and monitor the operating status of machines, detect faults and their causes early, and predict fault development trends. Its core lies in acquiring and analyzing the current state information of the equipment, which can include vibration signals, oil quality information, temperature conditions, ultrasonic or acoustic emission information, etc. As a closed-structure support component, the mechanical structure of hinge bearings makes it difficult to directly observe or detect internal faults. Therefore, bearing fault detection technology based on sensor information has been extensively studied.

[0003] In existing technologies, Lin Liangxing et al. proposed a method for predicting the remaining service life of coated spherical plain bearings based on VMD-EEMD-LSTM. They used variable mode decomposition and ensemble mode decomposition to extract features from the bearing friction torque signal, and then selected features based on temporal correlation. The hyperparameter optimization interval was chosen to perform Bayesian optimization on the LSTM, resulting in a Bayesian-optimized LSTM model. This model was used to predict the wear state of the spherical plain bearing, and experimental results showed high prediction accuracy and good generalization performance for bearings under different operating loads. Liu Yunfan et al. proposed a bearing remaining service life prediction model based on convolutional neural networks (CNN) and long short-term memory (LSTM) neural networks. First, CNN was used to extract failure features from the friction torque signal of the spherical plain bearing. Then, the torque signal, processed by principal component analysis (PCA) and filtering, was input into the LSTM neural network for training, resulting in a self-lubricating spherical plain bearing life prediction model that can accurately predict bearing wear. Results showed that coated self-lubricating spherical plain bearings can maintain a high reliability level (90%) and stable service for a long time under light load and low frequency conditions. Gao Yadong et al. set up different degrees of wear faults in joint bearings and measured the resulting airframe vibration signals. They then extracted the airframe vibration characteristics of these faults using spectral analysis. Using the spectral components of the fault signals as training and testing samples, and leveraging the good approximation capability of radial basis function neural networks, they achieved the ability to identify the wear degree of variable pitch tie rod joint bearings using only airframe vibration signals, with an average identification error of less than 10%. All of the above methods use deep learning to adaptively extract features, but they utilize the bearing's torque and vibration signals. For aerospace self-lubricating hinge bearings, torque signals are easily affected by temperature and environmental fluctuations, and vibration signals cannot accurately identify wear differences under low-angle, high-load conditions. Summary of the Invention

[0004] In view of the deficiencies in the prior art, the purpose of this invention is to provide a wear monitoring method and system for self-lubricating hinge bearings.

[0005] A wear monitoring method for a self-lubricating hinge bearing provided by the present invention includes:

[0006] Step S1: Simulate the real working scenario of the self-lubricating hinge bearing under constant working conditions. Based on the acoustic emission sensor and torque sensor installed on the outer ring of the self-lubricating hinge bearing and the coupling, collect the simulated voltage of the acoustic emission signal and torque signal during the swing process of the self-lubricating hinge bearing.

[0007] Step S2: Preprocess the analog voltage values ​​of the acquired acoustic emission signal and torque signal;

[0008] Step S3: Use empirical mode decomposition to perform noise reduction and feature extraction on the preprocessed analog voltage;

[0009] Step S4: Extract high-dimensional features that meet preset conditions using a separable convolutional neural network;

[0010] Step S5: Use a bidirectional gated neural network to identify the bearing wear state of high-dimensional features that meet the preset conditions, and classify the wear state according to the identified bearing wear state.

[0011] Preferably, step S1 includes connecting the acoustic emission sensor to the charge amplifier.

[0012] Preferably, step S2 involves: performing data outlier processing, missing data processing, and data time scale alignment processing on the analog voltage values ​​of the acquired acoustic emission and torque signals;

[0013] The outlier handling mentioned above is achieved by using the Z-score method to identify and process outliers.

[0014]

[0015] Where X represents a single data point; μ represents the mean; σ represents the standard deviation; a data point whose Z score exceeds the threshold is considered an anomaly.

[0016] The missing data handling method involves filling in the missing data using linear interpolation to maintain data integrity.

[0017]

[0018] Where x represents the time point of missing data, x1 and x2 are the time points before and after the missing point, and y1 and y2 are the corresponding data values;

[0019] The data timescale alignment process aligns all signal data on the time axis.

[0020] Preferably, step S3 involves: decomposing the preprocessed analog voltage quantity into multiple intrinsic mode functions (IMFs) using empirical mode decomposition; each IMF is an oscillating mode and satisfies two conditions, including: in the entire dataset, the number of extreme points and the number of zero crossover points must be equal or differ by at most one; and at any given time, the mean of the envelope defined by the local maxima and the envelope defined by the local minima is zero; based on the two conditions, IMFs are screened to obtain valid IMFs, invalid signal frequency bands are eliminated by screening valid IMFs, and valid features are retained; simultaneously, time-domain features, frequency-domain features, and time-frequency-domain features are extracted from the screened valid IMFs.

[0021] Preferably, the separable convolutional neural network includes: depthwise convolution and pointwise convolution;

[0022] The depthwise convolution is used to perform filtering operations on each input channel to extract spatial features;

[0023]

[0024] Where X represents the input feature map with M channels, and each convolutional kernel is K×K in size, then the convolutional kernel D for each channel m is... m Convolution operations are performed only on the corresponding input channels; Y m This represents the result of the m-th output channel;

[0025] The pointwise convolution is used to integrate the output of the depthwise convolution to achieve feature fusion;

[0026]

[0027] Pointwise convolution uses a 1×1 kernel to convolve the output of depthwise convolution; M indicates that the output after depthwise convolution has M channels; the pointwise convolution kernel has N output channels, and each output channel is processed by a 1×1 kernel that handles all input channels; where P n,m These are the weights of the 1×1 convolution kernel from the m-th input channel to the n-th output channel.

[0028] Preferably, the bidirectional gated loop unit network combines forward and reverse gated loop units to process the time series information corresponding to high-dimensional features; wherein, the gated loop unit optimizes the analog voltage of acoustic emission signals and torque signals containing high-dimensional features and time series information that meet preset requirements by updating and resetting gates.

[0029] Preferably, in the bidirectional gated loop unit network, the input of each time step is processed simultaneously by both forward and reverse gated loop units;

[0030] Update Gate Z t Control the previous state h t-1 Retain the current state h t Degree:

[0031] z t =σ(W z x t +U z h t-1 +b z )

[0032] Among them, W z U z It is the weight matrix, b z It is the bias term, σ is the sigmoid activation function; x t This represents the high-dimensional convolutional features of the input that meet the preset requirements;

[0033] Reset door r t This determines the extent to which previous state information is discarded:

[0034] r t =σ(W r x t +U r h t-1 +b r )

[0035] Similarly, using the weight matrix W r U r and bias b r ;

[0036] Candidate hidden state It is determined by both the current input and the past state adjusted by the reset gate:

[0037]

[0038] Where ⊙ represents element-wise product, W h U h It is the weight matrix, b h It is a bias term;

[0039] The final hidden state h t The fusion of the current candidate hidden state and the previous hidden state is regulated by the update gate:

[0040]

[0041] Based on the hidden state h tThe bearing wear condition is obtained.

[0042] A wear monitoring system for a self-lubricating hinge bearing according to the present invention includes:

[0043] Module M1: Simulates the real working scenario of a self-lubricating hinge bearing under constant working conditions. Based on the acoustic emission sensor and torque sensor installed on the outer ring of the self-lubricating hinge bearing and the coupling, it collects the simulated voltage of the acoustic emission signal and torque signal during the swing process of the self-lubricating hinge bearing.

[0044] Module M2: Preprocesses the analog voltage values ​​of the acquired acoustic emission and torque signals;

[0045] Module M3: Performs noise reduction and feature extraction on the preprocessed analog voltage using empirical mode decomposition;

[0046] Module M4: Extracted features utilize a separable convolutional neural network to extract high-dimensional features that meet preset conditions;

[0047] Module M5: Uses a bidirectional gated neural network to identify bearing wear status based on high-dimensional features that meet preset conditions, and classifies the wear status according to the identified bearing wear status.

[0048] Preferably, module M2 performs the following operations on the analog voltage values ​​of the acquired acoustic emission and torque signals: data outlier processing, data missing value processing, and data time scale alignment processing.

[0049] The outlier handling mentioned above is achieved by using the Z-score method to identify and process outliers.

[0050]

[0051] Where X represents a single data point; μ represents the mean; σ represents the standard deviation; a data point whose Z score exceeds the threshold is considered an anomaly.

[0052] The missing data handling method involves filling in the missing data using linear interpolation to maintain data integrity.

[0053]

[0054] Where x represents the time point of missing data, x1 and x2 are the time points before and after the missing point, and y1 and y2 are the corresponding data values;

[0055] The data timescale alignment process involves aligning all signal data along the time axis.

[0056] The module M3 employs the following approach: The preprocessed analog voltage quantity is decomposed into multiple intrinsic mode functions (IMFs) using empirical mode decomposition. Each IMF is an oscillating mode and satisfies two conditions: the number of extreme points and the number of zero-crossing points in the entire dataset must be equal or differ by at most one; simultaneously, at any given time, the mean of the envelopes defined by local maxima and local minima is zero. Based on these two conditions, valid IMFs are selected by filtering them. Invalid signal frequency bands are eliminated by filtering valid IMFs, while valid features are retained. Furthermore, time-domain features, frequency-domain features, and time-frequency-domain features are extracted from the selected valid IMFs.

[0057] Preferably, the separable convolutional neural network includes: depthwise convolution and pointwise convolution;

[0058] The depthwise convolution is used to perform filtering operations on each input channel to extract spatial features;

[0059]

[0060] Where X represents the input feature map with M channels, and each convolutional kernel is K×K in size, then the convolutional kernel D for each channel m is... m Convolution operations are performed only on the corresponding input channels; Y m This represents the result of the m-th output channel;

[0061] The pointwise convolution is used to integrate the output of the depthwise convolution to achieve feature fusion;

[0062]

[0063] Pointwise convolution uses a 1×1 kernel to convolve the output of depthwise convolution; M indicates that the output after depthwise convolution has M channels; the pointwise convolution kernel has N output channels, and each output channel is processed by a 1×1 kernel that handles all input channels; where P n,m These are the weights of the 1×1 convolution kernel from the m-th input channel to the n-th output channel;

[0064] The bidirectional gated cyclic unit network combines forward and reverse gated cyclic units to process time-series information corresponding to high-dimensional features; wherein, the gated cyclic unit optimizes the analog voltage of acoustic emission signals and torque signals containing high-dimensional features and time-series information that meet preset requirements by updating and resetting gates;

[0065] In the bidirectional gated loop unit network, the input at each time step is processed simultaneously by both forward and reverse gated loop units;

[0066] Update Gate Z tControl the previous state h t-1 Retain the current state h t Degree:

[0067] z t =σ(W z x t +U z h t-1 +b z )

[0068] Among them, W z U z It is the weight matrix, b z It is the bias term, σ is the sigmoid activation function; x t This represents the high-dimensional convolutional features of the input that meet the preset requirements;

[0069] Reset door r t This determines the extent to which previous state information is discarded:

[0070] r t =σ(W r x t +U r h t-1 +b r )

[0071] Similarly, using the weight matrix W r U r and bias b r ;

[0072] Candidate hidden state It is determined by both the current input and the past state adjusted by the reset gate:

[0073]

[0074] Where ⊙ represents element-wise product, W h U h It is the weight matrix, b h It is a bias term;

[0075] The final hidden state h t The fusion of the current candidate hidden state and the previous hidden state is regulated by the update gate:

[0076]

[0077] Based on the hidden state h t The bearing wear condition is obtained.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] 1. This invention avoids bearing wear detection while the machine is stopped, thereby realizing online condition monitoring and fault diagnosis;

[0080] 2. This invention uses both acoustic emission and torque sensor information, and extracts key wear state features by fusing features into a deep learning network. Compared with using torque or vibration information alone, it can more accurately and sensitively capture the differences between the initial stage of wear and low-speed conditions, and achieve higher precision wear state identification.

[0081] 3. This invention uses adaptive empirical mode decomposition to preprocess the data, reducing the problem of expert experience in selecting features, eliminating the interference of inappropriate features, and achieving data noise reduction while saving the cost of feature selection.

[0082] 4. This invention uses separable convolutional networks to extract features. By splitting standard convolution into depthwise convolution and pointwise convolution, the number of parameters and computational cost are significantly reduced. This not only improves computational efficiency but also reduces the risk of overfitting, making the model more suitable for mobile and embedded devices with limited resources. Attached Figure Description

[0083] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0084] Figure 1 This is a flowchart of the wear state identification method of the present invention.

[0085] Figure 2 This is a schematic diagram of torque signal acquisition data.

[0086] Figure 3 The waveforms of acoustic emission signals under different operating conditions are shown in the diagram.

[0087] Figure 4 This is a diagram of the integrated network structure.

[0088] Figure 5 This is a schematic diagram of a worn bearing outer ring gasket.

[0089] Figure 6 This is a schematic diagram of the worn inner ring of a bearing.

[0090] Figure 7 A scanning electron microscope image of the worn bearing surface.

[0091] Figure 8 This is a comparison chart of bearing wear accuracy between the present invention and other typical methods. Detailed Implementation

[0092] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0093] Example 1

[0094] This invention provides a wear monitoring method and system for self-lubricating hinge bearings. It is a multi-sensor fusion method and system for self-lubricating hinge bearings that can improve the real-time performance and accuracy of wear condition monitoring by integrating acoustic emission and torque.

[0095] The self-lubricating hinge bearing wear monitoring method, such as Figure 1 As shown, it includes:

[0096] Step 1: Mount the acoustic emission sensor on the outer ring of the self-lubricating spherical bearing. Install the torque sensor on the shaft to be tested via a coupling. Under constant working conditions, simulate the real working scenario. At the same time, connect the acoustic emission sensor to the charge amplifier to measure the simulated voltage of the acoustic emission signal and torque signal during the swing process of the self-lubricating hinge bearing.

[0097] In this embodiment, the acoustic emission sensor is connected to the charge amplifier to amplify the weak charge signal, improve the signal strength and quality, and achieve impedance matching for subsequent processing and analysis; this can effectively reduce noise and ensure the accuracy and stability of the signal.

[0098] Step 2: Perform data preprocessing on the collected signals, including handling outliers, handling missing values, and aligning the data time scale.

[0099] The preprocessing steps include identifying and handling outliers using the Z-score method to ensure that there are no extreme or non-standard values ​​in the dataset. The Z-score represents the distance of a data point from the mean, expressed as the standard deviation, as shown in the following formula:

[0100]

[0101] Where X is a single data point, μ is the mean, and σ is the standard deviation. A data point whose Z score exceeds a threshold, such as ±3σ, is considered an anomaly.

[0102] Missing information can be filled in using linear interpolation to maintain data integrity. Linear interpolation is suitable for data that is sequentially arranged with fixed time intervals. If the missing data is in a time series, the missing value can be estimated by connecting adjacent data points with a straight line, as shown in the following formula:

[0103]

[0104] Where x is the time point where the data is missing, x1 and x2 are the time points before and after the missing point, and y1 and y2 are the corresponding data values.

[0105] Finally, the alignment of the data time scales ensures the consistency of all signal data along the time axis, facilitating time series analysis and feature extraction. This series of data preprocessing operations forms the foundation for building efficient and reliable models.

[0106] Step 3: The preprocessed data is then denoised and feature extracted using Empirical Mode Decomposition (EMD). EMD aims to decompose complex data into a series of Intrinsic Mode Functions (IMFs), each of which is a simple oscillating mode and satisfies two fundamental conditions: first, the number of extreme points and the number of zero-crossing points in the entire dataset must be equal or differ by at most one; second, at any given time, the mean of the envelopes defined by local maxima and local minima is zero. Based on these two fundamental conditions, valid IMFs are selected. By selecting valid IMFs, invalid signal frequency bands are eliminated, retaining effective features. Furthermore, time-domain, frequency-domain, and time-frequency-domain features are extracted from the selected IMFs.

[0107] In this embodiment, the signal is segmented according to different wear states of the bearing (early wear stage: 0-10,000 cycles; stable wear stage: 10,000-30,000 cycles; rapid wear stage: 30,000-40,000 cycles; bearing failure stage: 40,000-50,000 cycles), and each segment is associated with its corresponding wear state. The data consists of the acoustic emission signal trigger value in the radial direction of the bearing outer ring and the bearing torque signal. Missing and outlier values ​​are processed for the signal data in each dimension, and empirical mode decomposition (EMF) is performed on the acoustic emission signal. Specifically, the highest frequency component in the signal is first identified, and the first intrinsic mode function (IMF) is extracted from it. Then, this IMF is subtracted from the original signal, and the next IMF is extracted. This process is repeated until all significant oscillation modes in the signal have been extracted or the remaining signal can no longer be decomposed into valid IMFs. The total length of the data in each dimension can be set to X. n Where n is the length of the data partition, after EMD decomposition, the data becomes X. m×n , where m is the number of groups in the IMF, and the one-dimensional data is transformed into two-dimensional data input model after empirical mode decomposition;

[0108] Step 4: Input the extracted features into a separable convolutional neural network. By designing separable convolutional kernels of different sizes, high-dimensional features are extracted.

[0109] Separable convolutional neural networks decompose the standard convolution operation into two smaller operations: depthwise convolution and pointwise convolution. First, depthwise convolution independently applies a separate kernel to each input channel. For each input channel, a corresponding kernel performs filtering only on that channel, thereby extracting spatial features. Pointwise convolution (i.e., 1x1 convolution) is used to integrate the outputs of depthwise convolution, mixing information across different channels to achieve feature fusion. Through these two steps, separable convolution effectively reduces the number of parameters and computational complexity, while still effectively extracting important features from the input data.

[0110] More specifically, depthwise convolution differs from standard convolution in feature extraction because it processes each channel individually instead of fusing information from multiple input channels. This significantly reduces the number of parameters and computational cost of convolutional neural networks. This approach enhances the computational efficiency of the model, making it particularly suitable for resource-constrained environments such as mobile devices and embedded systems. The formula is as follows:

[0111]

[0112] In depthwise convolution, each input channel is processed independently by a convolutional kernel. Assuming the input feature map is X with M channels, and each convolutional kernel is K×K in size, then the convolutional kernel D for each channel m... m (Size K×K) Convolution operations are performed only on the corresponding input channels. Where Y m This is the result of the m-th output channel.

[0113] Pointwise convolution uses kernels primarily for adjusting and fusing feature channels in convolutional neural networks. It can effectively change the dimension of channels, integrating features from different channels while reducing the number of parameters and computational cost. Through pointwise convolution, the network can maintain high performance while increasing nonlinearity and depth. The specific calculation formula is as follows:

[0114]

[0115] Pointwise convolution uses a 1×1 kernel to convolve the output of depthwise convolution, primarily used to combine features from different channels. If the output of depthwise convolution has M channels, the kernel of pointwise convolution has N output channels, with each output channel processed by a 1×1 kernel across all input channels. Where P... n,mThese are the weights of the 1×1 convolution kernel from the m-th input channel to the n-th output channel.

[0116] Step 5: The acoustic emission and torque signals generated during bearing wear have a temporal relationship. In order to explore the temporal variation patterns of relatively long intervals in the time series, a bidirectional gated neural unit is used to extract the time series features of the time series signal, and the data is transformed back into two dimensions. Finally, it is connected to the SoftMax layer to identify the bearing wear state and output the wear state classification.

[0117] Bidirectional gated neural networks (BNNs) employ special gating mechanisms to perform classification tasks. These gating structures control the inflow and outflow of information, thus addressing the vanishing or exploding gradient problems encountered by traditional recurrent neural networks (RNNs) in processing long sequences of data. In classification tasks, BNNs first process the sequence data through their recurrent structure, using input gates, forget gates, and output gates to update and maintain an internal state. This allows the network to capture long-term dependencies within the sequence. At each time step of the sequence, the network updates its state, and at the last time step, it uses one or more fully connected layers, typically coupled with activation functions such as softmax, to output a classification prediction.

[0118] More specifically, a Bi-directional Gated Recurrent Unit (Bi-GRU) is a special type of recurrent neural network that combines forward and backward GRUs to process sequential data, improving the ability to capture information. A GRU is a model that optimizes information flow through two gate mechanisms (update gate and reset gate).

[0119] In Bi-GRU, the input at each time step is processed simultaneously by both forward and backward GRUs, where:

[0120] Update Gate Z t Control the previous state h t-1 Retain the current state h t Degree:

[0121] z t =σ(W z x t +U z h t-1 +b z )

[0122] Among them, W z U z It is the weight matrix, b z σ is the bias term, and σ is the sigmoid activation function.

[0123] Reset door r tThis determines the extent to which previous state information is discarded:

[0124] r t =σ(W r x t +U r h t-1 +b r )

[0125] Similarly, using the weight matrix W r U r and bias b r .

[0126] Candidate hidden state It is determined by both the current input and the past state adjusted by the reset gate:

[0127]

[0128] Where ⊙ represents element-wise product, W h U h It is the weight matrix, b h It is a bias term.

[0129] The final hidden state h t The fusion of the current candidate hidden state and the previous hidden state is regulated by the update gate.

[0130]

[0131] Based on the hidden state h t The bearing wear condition is obtained.

[0132] Bi-GRU integrates contextual information at each time point through this mechanism, enabling it to capture the temporal dynamics of the sequence more comprehensively, thereby improving the model's performance in tasks such as speech recognition and text processing.

[0133] When the bearing is determined to be in the initial wear or normal wear stage, the hinge bearing can still work normally without the need for bearing replacement. When the bearing is determined to be in the rapid wear stage, the operator is reminded to be vigilant about wear and replacement of parts. When the bearing is determined to be in the failure stage, the operator is reminded to replace parts immediately to ensure the stable operation and safety of the equipment.

[0134] The present invention also provides a wear detection system for a self-lubricating hinge bearing. The wear detection system for the self-lubricating hinge bearing can be implemented by executing the process steps of the wear detection method for the self-lubricating hinge bearing. That is, those skilled in the art can understand the wear detection method for the self-lubricating hinge bearing as a preferred embodiment of the wear detection system for the self-lubricating hinge bearing.

[0135] Example 2

[0136] Example 2 is a preferred example of Example 1.

[0137] A wear monitoring method for a self-lubricating hinge bearing provided by the present invention includes:

[0138] The Qingcheng WM500-1 acoustic emission sensor was mounted on the outer ring of the self-lubricating spherical bearing. The acoustic emission sensor was connected to the data acquisition card SAEU3H-4 via a preamplifier SAEPA2. Simultaneously, the loading spindle was connected to the Lianyiyou DYN-200 torque sensor via a coupling. Then, under given operating conditions, using the cyclic swing angle and applied load, acoustic emission and torque signals were acquired. Figure 2 For torque data acquisition, three operating conditions are used: 1-no-load reciprocating oscillation (oscillation angle 17.5 degrees - 35 degrees), 2-no-load reciprocating oscillation (oscillation angle 35 degrees - 70 degrees), and 3-radial constant 3500N load oscillation (oscillation angle 17.5 degrees - 35 degrees). Figure 3 The acoustic emission signal acquisition data includes 10 given operating conditions: 1-no-load reciprocating oscillation (oscillation angle 17.5°-35°), 2-no-load reciprocating oscillation (oscillation angle 35°-70°), 3-constant radial load of 3500N (oscillation angle 17.5°-35°), 4-radial and axial fluctuating load (oscillation angle 17.5°-35°), and 5-constant radial load of 3500N and constant axial load of 350N (oscillation angle 17.5°-35°). 6 - Constant radial load of 3500N and constant axial load of 350N (swing angle 35-70 degrees), 7 - Constant radial load of 7000N and constant axial load of 350N (swing angle 35-70 degrees), 8 - Constant axial load of 350N (swing angle 17.5-35 degrees), 9 - Constant radial load of 3500N and constant axial load of 350N (swing angle 35-70 degrees), 10 - Constant axial load of 350N (swing angle 35-70 degrees).

[0139] After collecting sufficient data, data preprocessing is performed. Data in the same column is normalized to the range [0, 1], and outliers and missing values ​​are removed. To address the issue of noise and irrelevant components in the acoustic emission waveform, empirical mode decomposition (EMD) is used to process the signal, decomposing it into a series of intrinsic mode functions (EMFs). Irrelevant EMFs are filtered out, and the EMFs of the characteristic frequency bands are retained as data input. The data input format is [x1, x2, ..., x...]. n [,y], where x n Let y be the nth intrinsic mode function and y be the torque signal.

[0140] exist Figure 4The network structure shown comprises three main parts: First, data flows through a separable convolutional network, which consists of two stages: depthwise convolution and pointwise convolution. The pointwise convolution uses k kernels, which, compared to traditional convolution, maintains good feature extraction efficiency while reducing parameters. The extracted features are then input into a bidirectional gated recurrent neural network (Bi-GRU). The gating mechanism helps the model capture long-term and short-term dependencies in time series data. By finely adjusting the structure of the recurrent network, such as the number of units and layers, the model's sensitivity to temporal dynamics can be further improved. The final stage is a SoftMax classification layer, which converts the output of the recurrent network into a probability distribution of bearing wear states. This layer is simple and effective, clearly mapping the features learned by the network to specific wear state categories.

[0141] In separable convolution, the formula for separable convolution can be expressed as depthwise convolution DW-Conv(W,x). (i,j) And pointwise convolution PW-Conv(W,x) (i,j) The two operations are represented as follows:

[0142]

[0143] S-Conv(W D W P ,x) (i,j) =PW-Conv(W,x) (i,j) [W P ,DW-Conv(W,x) (i,j) ]

[0144] Where (i,j) is the pixel position in the output feature map, and (m,n) is the relative position in the convolution kernel g. The convolution kernel g typically slides on the input feature map f. For each position (i,j), the above double summation is calculated, where the convolution kernel element g(m,n) is multiplied by the corresponding element f(im,jn) in the input feature map, and these products are accumulated.

[0145] Using the output of a separable convolutional neural network as the input of a bidirectional gated neural network reduces the variance of the time series. The bidirectional gated recurrent neural network contains two gates: an update gate and a reset gate.

[0146] Update Gate Z t Control the previous state h t-1 Retain the current state h t The degree of [something] is calculated using the following formula:

[0147]

[0148] Among them W z U zIt is the weight matrix, b z σ is the bias term, and σ is the sigmoid activation function.

[0149] Reset door r t The determination of the extent to which previous state information is discarded is calculated using the following formula:

[0150]

[0151] Similarly, using the weight matrix W r U r and bias b r .

[0152] Candidate hidden state It is determined by both the current input and the past state adjusted by the reset gate, and the calculation formula is as follows:

[0153]

[0154] Where ⊙ represents element-wise product, W h U h It is the weight matrix, b h It is a bias term.

[0155] The final hidden state h t The fusion of the current candidate hidden state and the previous hidden state is regulated by the update gate, and the calculation formula is as follows:

[0156]

[0157] Bidirectional GRU integrates contextual information at each time point through this mechanism, enabling it to capture the temporal dynamics of the sequence more comprehensively, thereby improving the model's performance in wear and tear identification tasks.

[0158] The SoftMax layer is used for classification. The SoftMax function is a common activation function in machine learning for multi-class classification problems. It transforms the output of a vector into a probability distribution. For a specific classification task, if the original output or score (also called logits) produced by the last layer of the model is a vector Z = [z1, z2, ..., z...], then... K ], where K is the total number of categories, and the output σ(z) of the SoftMax function is a vector of length K, where each element σ(z) k This represents the probability estimate for the corresponding category. The function is defined as follows:

[0159]

[0160] Where z k It is the k-th element in vector z, corresponding to the score of the k-th class. It is zk The exponent, this transformation ensures that all output values ​​are positive. Denominator It is the sum of all index scores, used for normalization to ensure that the sum of all output values ​​is 1, thus forming an efficient probability distribution.

[0161] During backpropagation, the neural network updates its weights and biases using the gradient descent algorithm to minimize the loss function. The parameter update formula is expressed as:

[0162]

[0163] Where θ represents the network parameters (weights or biases). L is the loss function. It is the gradient of the loss function with respect to the parameters. η is the learning rate, which determines the update step size.

[0164] The model parameters are trained using the Adam optimizer, a widely used gradient descent method particularly suitable for handling large-scale and non-static optimization problems. It combines the ideas of momentum and RMSprop optimization techniques to achieve faster convergence and higher efficiency. The update rule of the Adam optimizer is as follows:

[0165] Calculate the exponential moving average of the gradient (first-moment estimate):

[0166]

[0167] Calculate the exponential moving average of the squared gradient (second moment estimate):

[0168]

[0169] Correct the estimates of the first and second moments (to prevent underestimation during the early stages of training):

[0170]

[0171] Parameter update:

[0172]

[0173] For the MS14101 self-lubricating hinge bearing, the acoustic emission signal sampling frequency was 10MHz, and the torque signal sampling frequency was 200Hz. After data segmentation, training and test sets were randomly divided. Wear data was recorded for 10 different loading conditions. Bearing wear can be classified into four different categories according to the degree of wear. In the neural network, these wear states are represented by four neurons in the output layer, with each neuron corresponding to one state. The activation of the output layer uses one-hot encoding, represented as: [1,0,0,0] (initial wear), [0,1,0,0] (normal wear stage), [0,0,1,0] (rapid wear stage), and [0,0,0,1] (failure stage). This encoding method ensures that each category is independent when the model performs classification, facilitating accurate identification of different wear states.

[0174] Figure 5 This is a diagram showing the wear of the bearing outer ring gasket. Figure 6 This is a partial schematic diagram after wire cutting of the inner ring of the bearing. Figure 7 The image shows a scanning electron microscope (SEM) image of the inner ring with a resolution of 50 micrometers. The image reveals slight ploughing abrasive wear and a certain degree of fatigue spalling wear on the inner ring of the bearing.

[0175] The network learning rate ε is set to 0.0005, gradient optimization is performed using the Adam method, and the maximum number of iterations is set to 100. Figure 8 To compare the wear accuracy of bearings under different operating conditions with other typical methods, the test hardware conditions were: AMD 5900X processor, 32G memory, and NVIDIA 3080 GPU.

[0176] This invention was validated using experimental data from 10 different working conditions. The experimental results were effective, and the wear stages were clearly defined. A Qingcheng WM500-1 acoustic emission sensor was mounted on the outer ring of a self-lubricating spherical bearing. The acoustic emission sensor was connected to a data acquisition card SAEU3H-4 via a preamplifier SAEPA2. Simultaneously, a loading mandrel was connected to a Lianyiyou DYN-200 torque sensor via a coupling. Data was collected under given working conditions. The data was normalized to the range [0,1] through normalization, outlier and missing value processing, and noise and irrelevant components were removed using adaptive mode decomposition. The preprocessed data was then input into the entire network structure. A separable convolutional neural network extracted high-dimensional features, and a bidirectional gated recurrent network searched for the correlation between the data and the wear stages. The method of simultaneously inputting the fused acoustic emission signal and torque signal into a separable convolutional network, as proposed in this invention, is innovative and feasible. Furthermore, it meets industrial requirements in terms of recognition accuracy and speed.

[0177] This invention demonstrates the feasibility of the method based on wear tests of self-lubricating hinge bearings under multiple working conditions. In the early and normal wear stages, the wear of the bearing does not affect the operating accuracy of the equipment, and there is no need to consider maintenance and replacement of parts. When it enters the rapid wear stage, it is necessary to prepare for replacement of parts and maintenance. When the model determines that the bearing has entered the failure stage, the operator needs to stop the machine in time and carry out maintenance and replacement of parts.

[0178] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0179] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for monitoring wear of a self-lubricating hinge bearing, characterized in that, include: Step S1: Simulate the real working scenario of the self-lubricating hinge bearing under constant working conditions. Based on the acoustic emission sensor and torque sensor installed on the outer ring of the self-lubricating hinge bearing and the coupling, collect the simulated voltage of the acoustic emission signal and torque signal during the swing process of the self-lubricating hinge bearing. Step S2: Preprocess the analog voltage values ​​of the acquired acoustic emission signal and torque signal; Step S3: Use empirical mode decomposition to perform noise reduction and feature extraction on the preprocessed analog voltage; Step S4: Extract high-dimensional features that meet preset conditions using a separable convolutional neural network; Step S5: Use a bidirectional gated neural network to identify the bearing wear state of high-dimensional features that meet the preset conditions, and classify the wear state according to the identified bearing wear state. Step S2 involves: processing outlier values, missing values, and aligning the time scale of the collected acoustic emission and torque signals to handle the analog voltage values. The outlier handling mentioned above is achieved by using the Z-score method to identify and process outliers. in, Represents a single data point; This represents the average value. Standard deviation; data points Scores exceeding the threshold are considered abnormal; The missing data handling method involves filling in the missing data using linear interpolation to maintain data integrity. in, Indicates the time point where data is missing. and These are the time points before and after the missing point. and These are the corresponding data values; The data timescale alignment process involves aligning all signal data along the time axis. The separable convolutional neural network includes: depthwise convolution and pointwise convolution; The depthwise convolution is used to perform filtering operations on each input channel to extract spatial features; in, Represents the input feature map, having There are channels, and the size of each convolutional kernel is . Then each channel convolution kernel Convolution operations are performed only on the corresponding input channels; Indicates the first The results of each output channel; The pointwise convolution is used to integrate the output of the depthwise convolution to achieve feature fusion; Pointwise convolution uses The convolution kernel performs convolution on the output of the depthwise convolution; The output after depthwise convolution has Each channel; the pointwise convolution kernel has There are 1 output channel, and each output channel consists of 1 The convolution kernel processes all input channels; among them, From the first The input channel to the first Each output channel The weights of the convolution kernel; In the bidirectional gated loop unit network, the input at each time step is processed simultaneously by both forward and reverse gated loop units; Update Gate Control the previous state Retain the current state Degree: in, It is a weight matrix. It is a bias term. It is the sigmoid activation function; This represents the high-dimensional convolutional features of the input that meet the preset requirements; Reset door This determines the extent to which previous state information is discarded: Similarly, a weight matrix is ​​used. and bias ; Candidate hidden state It is determined by both the current input and the past state adjusted by the reset gate: in, Represents element-wise product. It is a weight matrix. It is a bias term; Final hidden state The fusion of the current candidate hidden state and the previous hidden state is regulated by the update gate: ； Based on hidden state The bearing wear condition is obtained.

2. The wear monitoring method for self-lubricating hinge bearings according to claim 1, characterized in that, Step S1 includes connecting the acoustic emission sensor to the charge amplifier.

3. The wear monitoring method for self-lubricating hinge bearings according to claim 1, characterized in that, Step S3 involves: decomposing the preprocessed analog voltage quantity into multiple intrinsic mode functions (IMFs) using empirical mode decomposition; each IMF is an oscillating mode and satisfies two conditions, including: the number of extreme points and the number of zero crossover points in the entire dataset must be equal or differ by at most one; and at any given time, the mean of the envelope defined by the local maxima and the envelope defined by the local minima is zero; based on the two conditions, IMFs are screened to obtain valid IMFs, invalid signal frequency bands are eliminated by screening valid IMFs, and valid features are retained; simultaneously, time-domain features, frequency-domain features, and time-frequency-domain features are extracted from the screened valid IMFs.

4. The wear monitoring method for self-lubricating hinge bearings according to claim 1, characterized in that, The bidirectional gated loop unit network combines forward and reverse gated loop units to process time-series information corresponding to high-dimensional features; wherein, the gated loop unit optimizes the analog voltage of acoustic emission signals and torque signals containing high-dimensional features and time-series information that meet preset requirements by updating and resetting gates.

5. A wear monitoring system for a self-lubricating hinge bearing, characterized in that, include: Module M1: Simulates the real working scenario of a self-lubricating hinge bearing under constant working conditions. Based on the acoustic emission sensor and torque sensor installed on the outer ring of the self-lubricating hinge bearing and the coupling, it collects the simulated voltage of the acoustic emission signal and torque signal during the swing process of the self-lubricating hinge bearing. Module M2: Preprocesses the analog voltage values ​​of the acquired acoustic emission and torque signals; Module M3: Performs noise reduction and feature extraction on the preprocessed analog voltage using empirical mode decomposition; Module M4: Extracted features utilize a separable convolutional neural network to extract high-dimensional features that meet preset conditions; Module M5: Uses a bidirectional gated neural network to identify bearing wear status based on high-dimensional features that meet preset conditions, and classifies the wear status according to the identified bearing wear status. The module M2 performs the following operations on the analog voltage values ​​of the acquired acoustic emission and torque signals: data outlier processing, data missing value processing, and data time scale alignment processing. The outlier handling mentioned above is achieved by using the Z-score method to identify and process outliers. in, Represents a single data point; This represents the average value. Standard deviation; data points Scores exceeding the threshold are considered abnormal; The missing data handling method involves filling in the missing data using linear interpolation to maintain data integrity. in, Indicates the time point where data is missing. and These are the time points before and after the missing point. and These are the corresponding data values; The data timescale alignment process involves aligning all signal data along the time axis. The separable convolutional neural network includes: depthwise convolution and pointwise convolution; The depthwise convolution is used to perform filtering operations on each input channel to extract spatial features; in, Represents the input feature map, having There are channels, and the size of each convolutional kernel is . Then each channel convolution kernel Convolution operations are performed only on the corresponding input channels; Indicates the first The results of each output channel; The pointwise convolution is used to integrate the output of the depthwise convolution to achieve feature fusion; Pointwise convolution uses The convolution kernel performs convolution on the output of the depthwise convolution; The output after depthwise convolution has Each channel; the pointwise convolution kernel has There are 1 output channel, and each output channel consists of 1 The convolution kernel processes all input channels; among them, From the first The input channel to the first Each output channel The weights of the convolution kernel; In the bidirectional gated loop unit network, the input at each time step is processed simultaneously by both forward and reverse gated loop units; Update Gate Control the previous state Retain the current state Degree: in, It is a weight matrix. It is a bias term. It is the sigmoid activation function; This represents the high-dimensional convolutional features of the input that meet the preset requirements; Reset door This determines the extent to which previous state information is discarded: Similarly, a weight matrix is ​​used. and bias ; Candidate hidden state It is determined by both the current input and the past state adjusted by the reset gate: in, Represents element-wise product. It is a weight matrix. It is a bias term; Final hidden state The fusion of the current candidate hidden state and the previous hidden state is regulated by the update gate: ； Based on hidden state The bearing wear condition is obtained.

6. The wear monitoring system for self-lubricating hinge bearings according to claim 5, characterized in that, The module M3 employs the following approach: The preprocessed analog voltage quantity is decomposed into multiple intrinsic mode functions (IMFs) using empirical mode decomposition. Each IMF is an oscillating mode and satisfies two conditions: the number of extreme points and the number of zero-crossing points in the entire dataset must be equal or differ by at most one; simultaneously, at any given time, the mean of the envelopes defined by local maxima and local minima is zero. Based on these two conditions, valid IMFs are selected by filtering them. Invalid signal frequency bands are eliminated by filtering valid IMFs, while valid features are retained. Furthermore, time-domain features, frequency-domain features, and time-frequency-domain features are extracted from the selected valid IMFs.

7. The wear monitoring system for self-lubricating hinge bearings according to claim 5, characterized in that, The bidirectional gated loop unit network combines forward and reverse gated loop units to process time-series information corresponding to high-dimensional features; wherein, the gated loop unit optimizes the analog voltage of acoustic emission signals and torque signals containing high-dimensional features and time-series information that meet preset requirements by updating and resetting gates.

Images