A method for identifying the lubrication state of a self-lubricating spherical friction pair

Abstract

This invention provides a method and system for identifying the lubrication state of self-lubricating spherical friction pairs, belonging to the field of mechanical equipment condition monitoring. The method collects acoustic emission signals during the operation of the friction pair using an acoustic emission sensor. After normalization, overlapping sliding window sampling, and data augmentation preprocessing, the signals are input into a dual-stream time-frequency fusion deep learning model. The model extracts complementary features through parallel time-domain and frequency-domain feature extraction modules, performs spatiotemporal modeling using a bidirectional GRU and Transformer encoder layer, and optimizes the model by combining a focus loss function and the AdamW optimization algorithm. Finally, the lubrication state prediction result is output. This invention fully utilizes the time-frequency characteristics of acoustic emission signals, achieving high accuracy, strong anti-interference capability, and good real-time performance. It can effectively identify the lubrication break-in period, stable period, and rapid deterioration period, providing a guarantee for the reliable operation of self-lubricating spherical friction pairs.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment condition monitoring, and relates to a method for identifying the lubrication status of a self-lubricating spherical friction pair, specifically a method for monitoring the lubrication status of a self-lubricating spherical friction pair based on acoustic emission signals and a dual-stream time-frequency fusion network. Background Technology

[0002] Self-lubricating spherical friction pairs are core moving components in precision mechanical systems, widely used in key fields such as aerospace, precision manufacturing, and high-end equipment. During operation, they primarily rely on a transfer film formed at the friction interface by self-lubricating materials to reduce friction and wear. The lubrication state of the self-lubricating spherical friction pair directly affects the equipment's operating accuracy, transmission efficiency, service life, and operational safety. If lubrication degradation is not detected and addressed in a timely manner, it will lead to accelerated wear of the friction pair, abnormal temperature rise, seizing failure, and even significant economic losses and safety accidents.

[0003] Acoustic emission signals are high-frequency transient signals generated when materials release elastic energy during stress, friction, or damage. They can respond quickly to behaviors such as micro-friction, crack initiation, and material spalling. Therefore, they have the advantages of high sensitivity and strong real-time performance in the monitoring of friction pairs, making them an important monitoring signal source in this field.

[0004] Existing methods for monitoring lubrication status based on acoustic emission signals mostly employ a combination of manual feature extraction and traditional machine learning algorithms. Manual feature extraction is highly dependent on human experience and is difficult to fully characterize the complex nonlinear characteristics of acoustic emission signals. At the same time, most methods only use single-dimensional features in the time or frequency domains, resulting in insufficient information utilization and an inability to effectively model the dynamic process of lubrication status evolution over time. Furthermore, the accuracy and stability of identification are insufficient under complex working conditions and noise interference environments. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a method for identifying the lubrication state of self-lubricating spherical friction pairs. This method is based on a dual-stream deep learning network that integrates time-domain and frequency-domain features using acoustic emission technology to identify the lubrication state. This method has significant practical implications for extending the service life of self-lubricating spherical friction pairs and ensuring the reliable operation of related equipment. It can effectively solve the problems of composite domain offset and data scarcity in cross-regional assessment through a collaborative mechanism of global feature alignment and local instance correction, and significantly improve the prediction accuracy and spatial rationality of landslide susceptibility maps.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a method for identifying the lubrication state of a self-lubricating spherical friction pair, comprising the following steps: S1, Signal Acquisition and Preprocessing: Acoustic emission raw signals from different operating stages of the self-lubricated spherical friction pair are acquired, and the datasets corresponding to the lubrication break-in period, lubrication stabilization period, and rapid degradation period are divided and assigned category labels; global normalization processing is performed on the raw signals, and the processed long time-series signals are divided into fixed-length samples using an overlapping sliding window sampling method, and the number of samples is expanded through data augmentation; the dataset is divided into training set, validation set, and test set according to a preset ratio; S2. Construct a dual-stream time-frequency fusion deep learning model: Build a deep learning model with a parallel dual-stream feature extraction path that includes a time-domain feature extraction module, a frequency-domain feature extraction module, and a spatiotemporal feature fusion module. The parallel dual-stream structure effectively fuses time-domain and frequency-domain feature information, and the time-series modeling is used to mine the time dependence and global correlation of the signal. S3. Model Training and Optimization: The dual-stream time-frequency fusion deep learning model is trained using the training set. The model parameters are optimized and the optimal model is saved by adopting the focus loss function, AdamW optimization algorithm, cosine annealing dynamic learning rate scheduling strategy, mixed precision training and early stopping mechanism. S4. Lubrication status identification: After the acoustic emission signal to be identified is preprocessed in step S1, it is input into the optimal model, and the predicted probability distribution is output to determine the lubrication status of the self-lubricating spherical friction pair.

[0007] Preferably, in the process of acquiring the original acoustic emission signal in S1, an acoustic emission sensor arranged near the spherical friction pair is used to acquire the original acoustic emission signal throughout the entire operation. The acquired original acoustic emission signal is divided into different lubrication states according to the acoustic emission parameters. Different lubrication states correspond to different file directories. Multiple data files are stored in the directory. Each data file is in tabular format. The lubrication states include: lubrication break-in period, lubrication stabilization period, and rapid deterioration period. Different category labels are assigned to the above lubrication states.

[0008] Preferably, the global normalization process performed on the original signal in S1 specifically includes: performing a global normalization operation on the original acoustic emission signal, traversing all data, extracting the maximum and minimum values ​​in the entire dataset, and applying the global normalization formula: ; Process the signals to preserve the differences between different stages; in, The original signal value, The minimum value of the signal in the entire dataset. The maximum value of the signal in the entire dataset; The specific method of data augmentation is as follows: add Gaussian noise with an intensity of 5%-10% of the original signal amplitude, stretch the original signal to 110% of its original length or compress it to 90% of its original length, so as to effectively amplify a small number of samples and improve the balance of data distribution. The specific method for dividing the dataset is as follows: the balanced data is shuffled and merged into a total dataset, and then randomly divided into training set, validation set and test set according to a preset ratio of 7:1:2. Stratified sampling is used to maintain the proportion of each category during the division.

[0009] Preferably, the time-domain feature extraction module in S2 is used to extract local and global time-domain features from the time-domain waveform of the acoustic emission signal through multiple residual blocks. The specific extraction process is as follows: Using the time-domain waveform of the acoustic emission signal as input, local and global features are extracted through a feature extraction network consisting of three residual blocks; each residual block performs the following operations in sequence: Dilated convolution is used to expand the receptive field of the convolution kernel, capture long-range dependencies of signals, and extract local temporal features; After the convolution operation, a batch normalization layer is applied to normalize the input data; The ReLU activation function is used to introduce nonlinear mapping, which enhances the network's ability to express signal features; An embedded channel attention mechanism is used to enhance key temporal features by performing global average pooling on the features of each channel and learning channel weights in conjunction with a fully connected layer.

[0010] Preferably, the specific extraction process of the frequency domain feature extraction module in S2 is as follows: Perform a Fast Fourier Transform (FFT) on the input acoustic emission time-domain signal to convert it into a frequency-domain signal, thereby revealing the frequency components, periodic variations, and frequency characteristics of the signal. The converted frequency domain signal is input into a residual block network with the same structure as the time domain feature extraction module. The signal is then processed sequentially through dilated convolution, batch normalization, ReLU activation function, and channel attention mechanism to extract frequency domain features related to lubrication state.

[0011] Preferably, in S2, the spatiotemporal feature fusion module splices the extracted time-domain features and frequency-domain features to form a joint time-frequency feature sequence. The joint time-frequency features are then used for time-series modeling through a bidirectional gated recurrent unit (GRU) and a Transformer encoder layer to mine the signal's time dependence and global correlation. The specific process includes: The time-domain features output by the time-domain feature extraction module and the frequency-domain features output by the frequency-domain feature extraction module are concatenated and fused to generate a joint time-frequency feature sequence containing both time-domain and frequency-domain information, thereby achieving a comprehensive representation of signal features. The joint time-frequency feature sequence is input into the bidirectional gated cyclic unit (GRU) time series model. Through time series modeling, the temporal correlation and global association between features are mined, the time dependence in the time-frequency features is captured, and finally, a high-precision lubrication state prediction result is output.

[0012] Preferably, in S3, the early stopping mechanism is triggered when the performance of the validation set does not improve significantly within several consecutive epochs. At this point, training is stopped and the current optimal model parameters are saved.

[0013] Preferably, in S4, the evaluation metrics for lubrication status identification include overall accuracy, precision, recall, F1 score, and confusion matrix.

[0014] Preferably, in S2, the bidirectional GRU is used to capture the temporal dependence of time-frequency features, and the Transformer encoder layer is used to mine the global correlation between features. The two work together to improve the accuracy of state recognition.

[0015] Another aspect of the present invention provides a lubrication state identification system for a self-lubricating spherical friction pair, used to implement the method described above, the system comprising: The signal acquisition and preprocessing module is used to acquire the raw acoustic emission signals of the self-lubricating spherical friction pair in a time series, divide the datasets into different lubrication states, and perform global normalization, overlapping sliding window sampling, and data augmentation processing. The dual-stream time-frequency fusion model construction module is used to build a deep learning model that includes a time-domain feature extraction module, a frequency-domain feature extraction module, and a spatiotemporal feature fusion module. The time-domain feature extraction module and the frequency-domain feature extraction module both adopt a residual block structure with dilated convolution, batch normalization, ReLU activation function, and channel attention mechanism. The spatiotemporal feature fusion module fuses time-domain and frequency-domain features by splicing and fusing them, and performs time-series modeling through a bidirectional gated recurrent unit (GRU) and a Transformer encoder layer. The model training and optimization module is used to train the model using the preprocessed training dataset and optimize the model by combining the focus loss function, AdamW optimization algorithm, dynamic learning rate scheduling strategy, mixed precision training and early stopping mechanism. The lubrication state recognition module is used to input the preprocessed acoustic emission signal to be recognized into the trained optimal model, output the lubrication state prediction probability distribution, and complete the lubrication state recognition.

[0016] The present invention has the following beneficial effects: 1. This invention uses acoustic emission signals as the information source for characterizing the lubrication state of spherical friction pairs. By uniformly collecting and processing the acoustic emission signals generated during operation, it can effectively reflect the microscopic evolution characteristics of the contact interface of spherical friction pairs at different lubrication stages. Compared with traditional vibration or temperature signals, it has the advantages of fast response speed and high sensitivity.

[0017] 2. This invention improves sample utilization and enhances the ability to characterize local features by introducing global normalization processing and overlapping sliding window sampling, while ensuring that the physical properties of the signal are not destroyed. At the same time, combined with data augmentation methods such as noise perturbation and time scale transformation, the training sample size is effectively expanded, and the generalization ability of the model under complex working conditions and small sample conditions is significantly improved.

[0018] 3. This invention constructs parallel time-domain feature extraction paths and frequency-domain feature extraction paths, realizing collaborative modeling of time-domain and frequency-domain information of acoustic emission signals. This avoids the problem of insufficient information in a single feature domain, enabling the model to comprehensively characterize the lubrication state change features from multiple dimensions, thereby improving the accuracy and stability of lubrication state identification.

[0019] 4. This invention introduces residual structure and channel attention mechanism in the feature extraction process, which effectively alleviates the gradient vanishing problem in the deep network training process and enhances the model's ability to pay attention to key feature channels, thereby improving the ability to identify the differences in lubrication state features.

[0020] 5. In the model training process, this invention comprehensively adopts a loss function that considers class imbalance, an adaptive parameter optimization strategy, and a dynamic learning rate scheduling mechanism, which effectively improves the training efficiency and convergence stability of the model, reduces the risk of overfitting, and ensures the reliability of the recognition results. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This is a flowchart of the lubrication state identification method for spherical friction pairs according to the present invention.

[0023] Figure 2 This is a diagram of the dual-stream video fusion network structure of the present invention.

[0024] Figure 3 This is a working structure diagram of the residual block of the present invention.

[0025] Figure 4 This is a confusion matrix diagram according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the examples. However, those skilled in the art will understand that the following examples are only for illustrating this invention and should not be regarded as limiting the scope of this invention.

[0027] Example 1: See Figure 1 This embodiment provides a method for monitoring the lubrication status of spherical friction pairs based on acoustic emission signals and a dual-stream time-frequency fusion network, including the following steps: 1. Signal Acquisition and Preprocessing: Raw acoustic emission signals generated by the self-lubricated spherical friction pair at different operating stages were collected in a time series manner. Based on actual operating conditions, the collected acoustic emission signals were divided into datasets corresponding to different lubrication states. A unified global normalization process was performed on the raw acoustic emission signals to optimize model training. An overlapping sliding window sampling method was adopted to divide the long-time-series acoustic emission signals into fixed-length signal sample segments, with a preset overlap ratio between adjacent sample segments to preserve the temporal correlation of the signals and avoid the loss of key features. Data augmentation processing was performed on the signal sample segments by adding noise and time stretching to expand the number of samples, thereby effectively improving the model's generalization ability and anti-interference performance.

[0028] 2. Construct a dual-stream time-frequency fusion deep learning model: A deep learning model incorporating a parallel dual-stream feature extraction path is constructed. This model consists of a time-domain feature extraction module, a frequency-domain feature extraction module, and a spatiotemporal feature fusion module. The parallel dual-stream structure effectively fuses information from the time and frequency domains, and improves the accuracy of lubrication state prediction by mining the temporal dependence and global correlation of signals through time series modeling.

[0029] The purpose of the temporal feature extraction module is to extract local and global features from the temporal waveform of the acoustic emission signal. This module consists of three residual blocks (ResBlocks), each including the following key operations: Dilated convolution effectively captures long-range dependencies in the signal by expanding the receptive field of the convolutional kernel. The dilated convolutional layer extracts local features from the signal, helping the model capture the correlation between temporal signals with long strides. After each convolutional operation, a batch normalization layer normalizes the input data, ensuring the stability of network training and reducing gradient vanishing or exploding problems during training. The ReLU activation function is used after the convolutional layers, allowing the network to capture the nonlinear features of the input signal and improving the model's expressive power. The channel attention mechanism assigns different weights to each channel, helping the network focus on more important features. By performing global average pooling on each channel and combining it with fully connected layers to adjust the importance of channels, key signal features can be highlighted.

[0030] The purpose of the frequency domain feature extraction module is to extract spectral features related to lubrication state from the frequency domain components of the acoustic emission signal. The input time-domain signal is first converted into a frequency-domain signal using a Fast Fourier Transform (FFT). The FFT is used to extract the frequency components of the signal, helping to reveal the signal's periodic variations and frequency characteristics. The converted frequency-domain signal is then passed through the same improved residual block network as the time-domain feature extraction module to extract frequency-domain features. This network employs the same structure as the time-domain feature extraction module, including dilated convolutions, batch normalization, ReLU activation functions, and channel attention mechanisms. Through this process, the model is able to extract spectral features related to lubrication state from the frequency-domain signal.

[0031] The spatiotemporal feature fusion module fuses the features output from the time-domain feature extraction module and the frequency-domain feature extraction module to form a joint time-frequency feature sequence. It then uses time-series modeling to uncover the temporal correlations and global relationships between these features. This module fuses time-domain and frequency-domain features through concatenation, creating a joint time-frequency feature sequence. This joint feature sequence contains both time-domain and frequency-domain information, providing a more comprehensive signal representation. To uncover the temporal correlations and global relationships between features, the fused time-frequency feature sequence is processed by a time-series modeling network. A bidirectional gated recurrent unit (GRU) time-series model is used to model the joint time-frequency features. The goal of this time-series modeling network is to capture the temporal dependencies in the time-frequency features to further improve the prediction accuracy of lubrication status.

[0032] 3. Model Training and Optimization: The constructed dual-stream time-frequency fusion deep learning model was trained using a preprocessed training dataset. During training: to enhance the model's ability to identify difficult-to-classify samples and reduce the influence of easily-classified samples, this invention adopted a focus loss function. The focus loss function can improve the model's performance on imbalanced datasets by reducing the weight of easily-classified samples and focusing on difficult-to-classify samples. An adaptive optimization algorithm with weight decay, AdamW, was employed to avoid overfitting and accelerate convergence. A dynamic learning rate scheduling strategy was combined to adjust the learning rate during training. A mixed-precision training method was used to improve training efficiency and reduce computational resource consumption. An early stopping mechanism was introduced to save the optimal model parameters when the model performance no longer improves, thus preventing overfitting.

[0033] 4. Lubrication status identification: The acoustic emission signal from the spherical friction auxiliary channel to be identified is preprocessed in the same way and then input into the trained optimal dual-stream time-frequency fusion model. Based on the input signal, the model outputs the predicted probability distribution of each category.

[0034] Example 2: See Figure 1-4 This embodiment provides a method for identifying the lubrication state of a self-lubricating spherical friction pair, including: S1. Signal Acquisition and Preprocessing: S1.1 Acoustic Emission Signal Acquisition: Acoustic emission sensors positioned near the proximal end of the spherical friction pair are used to acquire raw acoustic emission signals throughout the entire operation. The acquired signals are categorized into different lubrication states based on acoustic emission parameters. Each lubrication state corresponds to a different file directory, storing multiple data files in tabular format. The lubrication states include: lubrication break-in period; lubrication stabilization period; and rapid deterioration period. Different category labels are assigned to each of these lubrication states.

[0035] S1.2 Acoustic emission signal preprocessing: The original acoustic emission signal is globally normalized. All data are traversed, and the maximum and minimum values ​​in the entire dataset are extracted and processed using formula (1) to retain the differences in signals at different stages and avoid feature loss caused by local normalization. The long time series signal is converted into a fixed-length sample by a sliding window. The window length L is 1024 sampling points, and the overlap rate is set to 75%. In order to improve the robustness and generalization ability of the model, the data is time-scaled. The scaling factor is 0.1, that is, the original signal is stretched to 110% of its original length to achieve effective amplification of a few samples and improve the balance of data distribution. The balanced data is shuffled and merged into a total dataset, and randomly divided into training set, validation set and test set according to the preset ratio of 7:1:2. Stratified sampling is used to maintain the proportion of each category during the division.

[0036] The formula for global normalization is: (1) in, The original signal value, The minimum value of the signal in the entire dataset. This represents the maximum value of the signal in the entire dataset.

[0037] S2. Construction of the time-frequency dual-stream feature extraction network: S2.1 Temporal Feature Extraction Module: The input is a 1×1024 one-dimensional sequence signal, processed through cascaded residual blocks. Each residual block introduces dilated convolutions to expand the receptive field without increasing the number of parameters. A dilation rate of 2 is set to capture long-range waveform features spanning multiple cycles in the acoustic emission signal. The residual blocks also include batch normalization and ReLU activation to enhance the network's nonlinear expressive power and facilitate signal feature learning. The first residual block increases the number of channels from 1 to 64, the second to 128, and the third to 256. Each residual block is followed by a max-pooling layer with a stride of 4 to reduce the temporal dimension of the signal, decrease computational cost, and retain the most important feature information. Finally, adaptive average pooling is used to compress the temporal dimension to a fixed length. Through adaptive pooling, input signals of different lengths can be converted into feature representations of the same dimension, facilitating subsequent processing.

[0038] S2.2 Frequency Domain Feature Extraction Module: Synchronously performs a one-dimensional Fast Fourier Transform on the input signal to extract its amplitude-frequency characteristics. Subsequent processing is completely symmetrical to the time-domain branch. This branch focuses on extracting the characteristic frequency distribution related to the acoustic emission signal under lubrication conditions of spherical friction, complementing the features of the time-domain branch.

[0039] S2.3 Spatiotemporal Feature Fusion Module: After feature extraction, the time-domain and frequency-domain features are concatenated and fused, and the fused time-frequency features are input into a bidirectional gated recurrent unit (GRU). The GRU can capture the evolution logic of the acoustic emission signal on the time axis and identify the dynamic trend of the lubrication state transitioning from the "break-in period" to the "rapid degradation period". The input dimension and hidden layer dimension of the GRU are both set to 512 to retain all feature information extracted by the two branches and avoid feature information loss due to dimensionality compression. Subsequently, a Transformer encoder layer is connected in series. Using a multi-head self-attention mechanism, the model can automatically identify the contrast relationship in the entire signal, significantly improving the ability to identify critical failure states. After completing the spatiotemporal modeling, global average pooling is used to aggregate the fused features to generate a fixed-size feature representation. Finally, the pooled features are used to perform state discrimination through a classifier containing two linear layers.

[0040] S3. Model Training and Optimization: S3.1 Focal Loss Function: To enhance the model's ability to identify difficult-to-classify samples and reduce the influence of easily-classified samples, a focal loss function is used. The focal loss function improves the model's performance on imbalanced datasets by reducing the weight of easily-classified samples and focusing on difficult-to-classify samples.

[0041] S3.2 Adaptive Optimization Algorithm with Weight Decay: In order to better update the model parameters, this invention adopts an adaptive optimization algorithm with weight decay (AdamW) to avoid overfitting and accelerate convergence.

[0042] S3.3 Dynamic Learning Rate Scheduling: To improve training efficiency and avoid excessively high or low learning rates during training, a dynamic learning rate scheduling strategy is adopted. Cosine annealing learning rate scheduling is used, which gradually reduces the learning rate during training, enabling the model to converge more stably and avoiding oscillations in the later stages of training.

[0043] S3.4 Early Stopping Mechanism: To prevent overfitting during training, this invention introduces an early stopping mechanism. During training, if the performance on the validation set does not show a significant improvement over several consecutive epochs, training is stopped and the parameters of the current optimal model are saved. This avoids excessively long training periods that could lead to the model overfitting the training data.

[0044] S4. Model Evaluation: Save the weights of the best-performing model during training. Perform a final evaluation on an independent test set, calculating the overall accuracy and confusion matrix (e.g., ...). Figure 4 (Example) and detailed metrics such as precision, recall, and F1 score. The loss and accuracy curves during training are recorded and plotted to analyze model convergence.

[0045] Example 3: This embodiment provides a lubrication status identification system for a self-lubricating spherical friction pair, the system comprising: The signal acquisition and preprocessing module is used to acquire the raw acoustic emission signals of the self-lubricating spherical friction pair in a time series, divide the datasets into different lubrication states, and perform global normalization, overlapping sliding window sampling, and data augmentation processing. The dual-stream time-frequency fusion model construction module is used to build a deep learning model that includes a time-domain feature extraction module, a frequency-domain feature extraction module, and a spatiotemporal feature fusion module. The time-domain feature extraction module and the frequency-domain feature extraction module both adopt a residual block structure with dilated convolution, batch normalization, ReLU activation function, and channel attention mechanism. The spatiotemporal feature fusion module fuses time-domain and frequency-domain features by splicing and fusing them, and performs time-series modeling through a bidirectional gated recurrent unit (GRU) and a Transformer encoder layer. The model training and optimization module is used to train the model using the preprocessed training dataset and optimize the model by combining the focus loss function, AdamW optimization algorithm, dynamic learning rate scheduling strategy, mixed precision training and early stopping mechanism. The lubrication state recognition module is used to input the preprocessed acoustic emission signal to be recognized into the trained optimal model, output the lubrication state prediction probability distribution, and complete the lubrication state recognition.

[0046] Furthermore, in the signal acquisition and preprocessing module, the global normalization formula is: ; in, The original signal value, The minimum value of the signal in the entire dataset. The maximum value of the entire dataset signal is given; the window length of the overlapping sliding window is 1024 sampling points, and the overlap rate is 75%; data augmentation processing includes adding noise and time stretching, with a time stretching scaling factor of 0.1.

[0047] Furthermore, the temporal feature extraction module includes three cascaded residual blocks. The first residual block increases the number of channels from 1 to 64, the second to 128, and the third to 256. Each residual block is followed by a max pooling layer with a stride of 4, and finally, the temporal dimension is compressed to a fixed length through adaptive average pooling. The dilated convolutional dilation rate of the residual block is 2.

[0048] Furthermore, the frequency domain feature extraction module first converts the time domain signal into a frequency domain signal using a fast Fourier transform (FFT) and extracts the amplitude-frequency characteristics. The subsequent processing structure is completely consistent with that of the time domain feature extraction module, thereby realizing frequency domain feature extraction.

[0049] Furthermore, in the spatiotemporal feature fusion module, the input dimension and hidden layer dimension of the bidirectional GRU are both set to 512; the Transformer encoder layer mines the global correlation of features through a multi-head self-attention mechanism; the classifier consists of two linear layers and is used to output the lubrication state prediction result.

[0050] Although the preferred embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many specific modifications under the guidance of the present invention without departing from the spirit of the invention and the scope of protection of the claims, and these modifications all fall within the scope of protection of the present invention.

Claims

1. A method for identifying the lubrication state of a self-lubricating spherical friction pair, characterized in that, Includes the following steps: S1, Signal Acquisition and Preprocessing: Acoustic emission raw signals from different operating stages of the self-lubricated spherical friction pair are acquired, and the datasets corresponding to the lubrication break-in period, lubrication stabilization period, and rapid degradation period are divided and assigned category labels; global normalization processing is performed on the raw signals, and the processed long time-series signals are divided into fixed-length samples using an overlapping sliding window sampling method, and the number of samples is expanded through data augmentation; the dataset is divided into training set, validation set, and test set according to a preset ratio; S2. Construct a dual-stream time-frequency fusion deep learning model: Build a deep learning model with a parallel dual-stream feature extraction path that includes a time-domain feature extraction module, a frequency-domain feature extraction module, and a spatiotemporal feature fusion module. The parallel dual-stream structure effectively fuses time-domain and frequency-domain feature information, and the time-series modeling is used to mine the time dependence and global correlation of the signal. S3. Model Training and Optimization: The dual-stream time-frequency fusion deep learning model is trained using the training set. The model parameters are optimized and the optimal model is saved by adopting the focus loss function, AdamW optimization algorithm, cosine annealing dynamic learning rate scheduling strategy, mixed precision training and early stopping mechanism. S4. Lubrication status identification: After the acoustic emission signal to be identified is preprocessed in step S1, it is input into the optimal model, and the predicted probability distribution is output to determine the lubrication status of the self-lubricating spherical friction pair.

2. The method for identifying the lubrication state of a self-lubricating spherical friction pair according to claim 1, characterized in that, In the process of acquiring the original acoustic emission signal in S1, the acoustic emission sensor arranged near the spherical friction pair is used to acquire the original acoustic emission signal throughout the entire operation. The acquired original acoustic emission signal is divided into different lubrication states according to the acoustic emission parameters. Different lubrication states correspond to different file directories. Multiple data files are stored in the directory. Each data file is in tabular format. The lubrication states include: lubrication break-in period, lubrication stabilization period, and rapid deterioration period. Different category labels are assigned to the above lubrication states.

3. The method for identifying the lubrication state of a self-lubricating spherical friction pair according to claim 1, characterized in that, The global normalization process for the original signal in S1 specifically includes: performing a global normalization operation on the original acoustic emission signal, traversing all data, extracting the maximum and minimum values ​​from the entire dataset, and applying the global normalization formula: ； Process the signals to preserve the differences between different stages; in, The original signal value, The minimum value of the signal in the entire dataset. The maximum value of the signal in the entire dataset; The specific method of data augmentation is as follows: add Gaussian noise with an intensity of 5%-10% of the original signal amplitude, stretch the original signal to 110% of its original length or compress it to 90% of its original length, so as to effectively amplify a small number of samples and improve the balance of data distribution. The specific method for dividing the dataset is as follows: the balanced data is shuffled and merged into a total dataset, and then randomly divided into training set, validation set and test set according to a preset ratio of 7:1:

2. Stratified sampling is used to maintain the proportion of each category during the division.

4. The method for identifying the lubrication state of a self-lubricating spherical friction pair according to claim 1, characterized in that, The time-domain feature extraction module in S2 is used to extract local and global time-domain features from the time-domain waveform of the acoustic emission signal through multiple residual blocks. The specific extraction process is as follows: Using the time-domain waveform of the acoustic emission signal as input, local and global features are extracted through a feature extraction network consisting of three residual blocks; each residual block performs the following operations in sequence: Dilated convolution is used to expand the receptive field of the convolution kernel, capture long-range dependencies of signals, and extract local temporal features; After the convolution operation, a batch normalization layer is applied to normalize the input data; The ReLU activation function is used to introduce nonlinear mapping, which enhances the network's ability to express signal features; An embedded channel attention mechanism is used to enhance key temporal features by performing global average pooling on the features of each channel and learning channel weights in conjunction with a fully connected layer.

5. The method for identifying the lubrication state of a self-lubricating spherical friction pair according to claim 4, characterized in that, The specific extraction process of the frequency domain feature extraction module in S2 is as follows: Perform a Fast Fourier Transform (FFT) on the input acoustic emission time-domain signal to convert it into a frequency-domain signal, thereby revealing the frequency components, periodic changes, and frequency characteristics of the signal. The converted frequency domain signal is input into a residual block network with the same structure as the time domain feature extraction module. The signal is then processed sequentially through dilated convolution, batch normalization, ReLU activation function, and channel attention mechanism to extract frequency domain features related to lubrication state.

6. The method for identifying the lubrication state of a self-lubricating spherical friction pair according to claim 5, characterized in that, The spatiotemporal feature fusion module in S2 splices the extracted time-domain features and frequency-domain features to form a joint time-frequency feature sequence. The joint time-frequency features are then used for time-series modeling through a bidirectional gated recurrent unit (GRU) and a Transformer encoder layer to mine the time dependence and global correlation of the signal. The specific process includes: The time-domain features output by the time-domain feature extraction module and the frequency-domain features output by the frequency-domain feature extraction module are concatenated and fused to generate a joint time-frequency feature sequence containing both time-domain and frequency-domain information, thereby achieving a comprehensive representation of signal features. The joint time-frequency feature sequence is input into the bidirectional gated cyclic unit (GRU) time series model. Through time series modeling, the temporal correlation and global association between features are mined, the time dependence in the time-frequency features is captured, and finally, a high-precision lubrication state prediction result is output.

7. The method for identifying the lubrication state of a self-lubricating spherical friction pair according to claim 1, characterized in that, In S3, the early stopping mechanism is triggered when the validation set performance does not improve significantly within several consecutive epochs. At this point, training is stopped and the current optimal model parameters are saved.

8. The method for identifying the lubrication state of a self-lubricating spherical friction pair according to claim 1, characterized in that, In S4, the evaluation metrics for lubrication status identification include overall accuracy, precision, recall, F1 score, and confusion matrix.

9. The method for identifying the lubrication state of a self-lubricating spherical friction pair according to claim 6, characterized in that, In S2, the bidirectional GRU is used to capture the temporal dependence of time-frequency features, and the Transformer encoder layer is used to mine the global correlation between features. The two work together to improve the accuracy of state recognition.

10. A lubrication state identification system for a self-lubricating spherical friction pair, used to implement the method according to any one of claims 1 to 9, characterized in that, The system includes: The signal acquisition and preprocessing module is used to acquire the raw acoustic emission signals of the self-lubricating spherical friction pair in a time series, divide the datasets into different lubrication states, and perform global normalization, overlapping sliding window sampling, and data augmentation processing. The dual-stream time-frequency fusion model construction module is used to build a deep learning model that includes a time-domain feature extraction module, a frequency-domain feature extraction module, and a spatiotemporal feature fusion module. The time-domain feature extraction module and the frequency-domain feature extraction module both adopt a residual block structure with dilated convolution, batch normalization, ReLU activation function, and channel attention mechanism. The spatiotemporal feature fusion module fuses time-domain and frequency-domain features by splicing and fusing them, and performs time-series modeling through a bidirectional gated recurrent unit (GRU) and a Transformer encoder layer. The model training and optimization module is used to train the model using the preprocessed training dataset and optimize the model by combining the focus loss function, AdamW optimization algorithm, dynamic learning rate scheduling strategy, mixed precision training and early stopping mechanism. The lubrication state recognition module is used to input the preprocessed acoustic emission signal to be recognized into the trained optimal model, output the lubrication state prediction probability distribution, and complete the lubrication state recognition.

Images