A lightweight model-based fault diagnosis method for aircraft engine rolling bearings

Abstract

The present invention provides a method for diagnosing rolling bearing faults in an aircraft engine using a lightweight model, comprising: first, in a signal acquisition step, converting a vibration signal into a two-dimensional time-frequency image; then, inputting this two-dimensional image into an improved RegNet network for processing to achieve fault classification. The network consists of an input backbone, a main body, and an output head. The main body includes multiple stages, each stage having several stacked blocks, each block including multiple convolutional layers and an ELU activation layer. A channel-multiscale spatial attention mechanism (CMSAM) is applied before the last convolutional layer to enhance feature extraction capabilities. The CMSAM attention mechanism consists of two parts: channel attention and multiscale spatial attention, which enhance the feature extraction capabilities of RegNet. The technical solution of the present invention significantly reduces computational complexity and resource requirements, while improving the robustness and fault identification capabilities of the model in high-noise environments.

Description

Technical Field

[0001] The invention belongs to the technical field of fault diagnosis and relates to a method for diagnosing rolling bearing faults of a lightweight model of an aero-engine. Background Art

[0002] Aircraft engines are core components of aircraft, and their stable operation ensures safe flight. Rolling bearings are crucial load-bearing components within aircraft engines, and are also among the most susceptible to damage. Their performance directly impacts the service life and reliability of the aircraft engine. Aircraft engine rolling bearings often operate under harsh conditions, high loads, and variable operating conditions for extended periods, which can easily lead to performance degradation and the inevitable occurrence of various faults such as pitting, wear, and shedding. Failure to promptly repair or replace faulty bearings can directly cause aircraft equipment downtime, reduce aircraft safety, cause economic losses, and even lead to devastating damage and catastrophic accidents. Therefore, research on aircraft engine bearing fault diagnosis is crucial to promptly identify the early signs of minor aircraft rolling bearing faults and improve the reliability and safety of aircraft engine systems.

[0003] Over the past few years, aircraft engine bearing fault diagnosis technology has developed rapidly. Deep learning technologies, such as convolutional neural networks (CNNs), have become a mainstream approach to fault diagnosis. With their powerful feature extraction and pattern recognition capabilities, they are widely used in mechanical fault diagnosis fields such as aerospace, wind power generation, and petrochemicals. Although CNN methods demonstrate good adaptability and generalization in identification tasks, enhancing the model's focus on important fault characteristics and improving the effectiveness of rolling bearing fault diagnosis under complex operating conditions, the overall design approach still has some unresolved flaws:

[0004] 1) Deep CNNs have enhanced their feature learning capabilities to a certain extent. However, as the network size increases, its structure becomes increasingly complex, resulting in a large number of model parameters, requiring a large number of training samples and training times. This leads to high training time and computational overhead, resulting in high computational complexity and low fault identification efficiency. To improve the model's feature extraction capabilities, some methods have introduced computationally intensive attention mechanisms, such as multi-scale convolution modules. However, multi-scale convolutional layers introduce a large number of convolution operations, making the model complex and large in scale, resulting in suboptimal efficiency.

[0005] 2) Considering that some application scenarios need to be ported to embedded or mobile devices with limited hardware resources, the above models still face many constraints, such as many model parameters, long training time, and large storage space requirements. Therefore, there is an urgent need to study more lightweight and efficient deep learning models and other artificial intelligence technologies for application in the field of fault diagnosis.

[0006] 3) Each convolution in the convolutional neural network is followed by a ReLU activation function, which is an activation function with relatively high overhead.

[0007] Therefore, how to provide an aircraft engine rolling bearing fault diagnosis method that can ensure fault identification efficiency while meeting lightweight requirements is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] In response to the problems existing in the above work, such as the large number of parameters, high complexity, and high computational overhead of the improved CNN model leading to low efficiency, and the weak feature extraction of the proposed lightweight model, this paper proposes a new attention mechanism (CMSAM) and a fault diagnosis method for bearings based on the RegNet-CMSAM lightweight model.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] The present invention discloses a method for diagnosing rolling bearing faults of an aero-engine using a lightweight model, comprising the following steps:

[0011] Signal acquisition steps:

[0012] Collecting vibration signals during the operation of an aircraft engine rolling bearing and converting the vibration signals into a two-dimensional time-frequency image;

[0013] Improved RegNet network troubleshooting steps:

[0014] The two-dimensional time-frequency image is input into the improved RegNet network; the improved RegNet network sequentially includes an input backbone network, a main network, and an output head network; the main network includes several stages, each stage includes several stacked blocks, each block sequentially includes n convolutional layers and an Elu activation layer, and a channel-multi-scale spatial attention mechanism is applied before the nth convolutional layer for feature extraction, where n>2; the two-dimensional time-frequency image is sequentially input into the several stages through the input backbone network to obtain feature results, and the output head network receives the feature results, establishes a mapping relationship between the fault type and the feature results, and implements fault classification;

[0015] Channel-multi-scale spatial attention mechanism feature extraction steps:

[0016] Based on the channel attention mechanism, the feature map output by the n-1th convolutional layer is extracted to obtain the channel attention weighted feature. Figure 1 ; Based on the multi-scale spatial attention mechanism, the features Figure 1 Perform feature extraction to obtain features weighted for spatial channels Figure 2 ; The characteristics Figure 2 Input to the nth convolutional layer.

[0017] Preferably, the signal acquisition step includes: using continuous wavelet transform to convert the vibration signal into a two-dimensional time-frequency image

[0018] Preferably, the complex Morlet wavelet is used as the mother wavelet of the continuous wavelet transform.

[0019] Preferably, the method further includes improving the training steps of the RegNet network:

[0020] Collecting vibration signals of rolling bearings in different failure modes during operation of aircraft engines as training data; the failure modes include normal state, inner race failure, and outer race failure;

[0021] Using the training data to train the improved RegNet network to obtain a trained improved RegNet network;

[0022] The trained improved RegNet network is used to perform a fault diagnosis step.

[0023] Preferably, the step of extracting features from the feature map output by the n-1th convolutional layer based on the channel attention mechanism includes:

[0024] Through adaptive global average pooling, the feature map output by the n-1th convolutional layer is compressed in the spatial dimension to obtain the compressed features;

[0025] The compressed features are restored to the original number of channels through convolution, and the channel features are restored and integrated;

[0026] The Sigmoid activation function is used to normalize the weight value of each channel to the range of [0, 1]. The generated weight is multiplied channel by channel with the feature map output by the n-1th convolutional layer, and finally the feature map after channel attention weighting is obtained.

[0027] Preferably, the features are analyzed based on a multi-scale spatial attention mechanism. Figure 1 The steps for feature extraction include:

[0028] Use multi-scale depth-wise separable convolution to extract spatial attention features;

[0029] Add the spatial attention features extracted at different scales to obtain multi-scale spatial attention weight values;

[0030] The obtained weights are compared with the features Figure 1 Multiply channel by channel to obtain the weighted features of the spatial channels Figure 2 .

[0031] Preferably, the depthwise separable convolution operation at each scale includes the following steps:

[0032] For the characteristics Figure 1 Performing a depth convolution operation, including: performing a convolution operation on each channel using an independent convolution kernel to extract deep features;

[0033] Performing a point convolution operation on the depth feature includes: using a 1×1 convolution kernel to perform a convolution operation on the depth feature to fuse information between channels.

[0034] Preferably, in the subject network:

[0035] Each stage's Block includes a main branch and a shortcut branch;

[0036] The main branch adopts group convolution operation;

[0037] The shortcut branch adds the input of the main branch to the output of the last layer of the main branch.

[0038] Preferably, the main network includes 4 stages, namely stage 1, stage 2, stage 3 and stage 4;

[0039] The Block shortcut branch in stage 1 is not processed, and the resolution of input and output remains consistent;

[0040] The Block shortcut branches in stage 2, stage 3, and stage 4 perform downsampling processing, and the output resolution is reduced exponentially compared to the input resolution.

[0041] Preferably, the number of stage2, stage3 and stage4 increases sequentially.

[0042] As can be seen from the above technical solution, compared with the existing method of relying on predefined reasoning paths to achieve diagnosis through simple search and matching, the beneficial effects of the present invention include:

[0043] (1) By combining an improved RegNet lightweight model with the Channel-Multiscale Spatial Attention Mechanism (CMSAM), this paper effectively reduces the computational complexity of the model while improving the accuracy and robustness of fault diagnosis. Experimental results show that the fault identification accuracy of this method reaches 99.25%.

[0044] (2) This invention uses a RegNet model optimized for the task of aircraft engine rolling bearing fault diagnosis. This improves the original network structure and introduces the more efficient ELU activation function instead of the original ReLU activation function. This reduces the number of parameters and computational complexity, enhances the model's deployability in resource-limited environments, and adapts to the needs of efficient real-time diagnosis. While maintaining high diagnostic performance, the model's number of parameters and floating-point computational complexity are significantly reduced. Compared with other advanced models, this invention has lower parameter and computational complexity.

[0045] (3) A new channel-multiscale spatial attention mechanism is proposed, which can effectively extract multi-level and multi-scale feature information, thereby improving the recognition ability of different fault types. Compared with existing attention methods (such as SE, CBAM, GAM, etc.), this mechanism performs better in terms of accuracy, computational efficiency and noise resistance.

[0046] (4) The robustness of the proposed method in different noise environments was verified by experiments on HIT and DIGR datasets, indicating that it can maintain high diagnostic accuracy under noise conditions of -2dB ~ 4dB, and is particularly suitable for real-time health monitoring and fault warning of rolling bearings in aircraft engines.

[0047] (5) The diagnostic method of the present invention can still maintain a high diagnostic accuracy under high speed conditions. At high speeds in the DIGR dataset, the accuracy rate reaches 99.25%, demonstrating its adaptability and stability to high dynamic conditions in practical applications.

[0048] (6) The present invention combines lightweight model design with real-time processing requirements, not only maintaining high accuracy but also significantly reducing computational complexity and storage requirements, ensuring that the model can be successfully deployed in resource-constrained embedded devices or real-time monitoring systems, and is suitable for fault diagnosis and early warning in actual industrial environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only embodiments of the present invention. Those skilled in the art can also derive other drawings based on the provided drawings without inventive effort.

[0050] Figure 1 A schematic diagram of a method for diagnosing rolling bearing faults in an aircraft engine using a lightweight model according to an embodiment of the present invention;

[0051] Figure 2Schematic diagram of the RegNet-CMSAM model diagnostic process provided by an embodiment of the present invention;

[0052] Figure 3 Schematic diagram of CWT time-frequency images of seven types of faults provided by an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of the CMSAM structure provided in an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of the RegNet network architecture provided in an embodiment of the present invention;

[0055] Figure 6 t-SNE visualization results of the RegNet-CMSAM model test results provided by an embodiment of the present invention;

[0056] Figure 7 Schematic diagram of the confusion matrix of the RegNet-CMSAM model test results provided by an embodiment of the present invention;

[0057] Figure 8 A schematic diagram of the fault recognition accuracy of the five methods provided in the embodiments of the present invention under different signal-to-noise ratio conditions. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] This embodiment of the present invention provides a novel method for aircraft engine rolling bearing fault diagnosis using a channel-based multi-scale attention mechanism (CMSAM) and an improved lightweight RegNet network model (RegNet-CMSAM). First, the vibration signal is converted into a two-dimensional time-frequency image using a continuous wavelet transform. Then, the CMSAM is designed and combined with a modified RegNet model using the ELU activation function for fault identification and classification.

[0060] like Figure 1 As shown, the following steps are included:

[0061] Signal acquisition steps:

[0062] Collect vibration signals from rolling bearings in aircraft engines during operation and convert them into two-dimensional time-frequency images;

[0063] Improved RegNet network troubleshooting steps:

[0064] The two-dimensional time-frequency image is input into the improved RegNet network; the improved RegNet network sequentially includes an input backbone network, a main network, and an output head network; the main network includes several stages, each stage includes several stacked blocks, each block includes n convolutional layers and an Elu activation layer, and a channel-multi-scale spatial attention mechanism is applied before the nth convolutional layer for feature extraction, where n>2; the two-dimensional time-frequency image is sequentially input into several stages through the input backbone network to obtain feature results, and the output head network receives the feature results, establishes a mapping relationship between fault type and feature results, and realizes fault classification;

[0065] Channel-multi-scale spatial attention mechanism feature extraction steps:

[0066] Based on the channel attention mechanism, the feature map output by the n-1th convolutional layer is extracted to obtain the channel attention weighted feature. Figure 1 ; Based on the multi-scale spatial attention mechanism for features Figure 1 Perform feature extraction to obtain features weighted for spatial channels Figure 2 ;feature Figure 2 Input to the nth convolutional layer. A channel-multi-scale spatial attention mechanism (CMSAM) is added before the last convolutional layer of each stage of the improved RegNet network, allowing the model to pay more attention to important channel features, thereby enhancing the feature extraction capability of RegNet.

[0067] In one embodiment, before performing the two-dimensional time-frequency image conversion, a vibration signal normalization step is also included:

[0068] The original data is a one-dimensional vibration signal time series. In order to eliminate the scale differences between different signal features and facilitate subsequent model training and feature learning, we standardize the original data. The purpose of standardization is to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby improving the convergence speed and stability of the model. Specifically, assuming that the total number of samples in the data set is n, the vibration signal data of each sample is represented by x i (where i = 1, 2, ..., n). Each sample is processed using the Z-score normalization method, which is calculated as follows:

[0069] ;

[0070] Among them, Z(x i ) is the standardized data; x i is the vibration signal data; μ is the average value of the signal data; is the standard deviation of the signal data. μ and The calculation formulas are:

[0071] ;

[0072] ;

[0073] Through the above processing, the standardized data has zero mean and unit variance, which can eliminate the deviation caused by amplitude differences in the signal data, so that subsequent deep learning models or other diagnostic algorithms can more effectively process signal features, reduce the impact of different data scales, and improve diagnostic performance.

[0074] In one embodiment, because it is difficult to distinguish obvious fault characteristics from the time or frequency domain information of the vibration signal, the combination of time and frequency domains provides more valuable information for the rolling bearing fault diagnosis method based on deep learning. Therefore, a continuous wavelet transform is introduced to extract the time and frequency domain features of the signal.

[0075] The continuous wavelet transform (CWT) is a time-frequency analysis method that obtains the characteristics of a signal at different times and frequencies by applying wavelet functions of different scales to the signal. The basic principle is to perform a wavelet transform on the input signal and extract local features using wavelet basis functions with adjustable scales. Given a square-integrable signal f(t), the calculation formula of the CWT is defined as:

[0076] ;

[0077] Among them, f(x) is the original signal, t is time, α is the stretching factor, τ is the translation factor, is the wavelet basis function, yes The complex conjugate of .

[0078] In this embodiment, selecting a suitable wavelet basis function is the key to effectively performing wavelet transform. Since the impact characteristics of the collected bearing vibration signal are similar to those of the Morlet wavelet, and the complex Morlet wavelet is the complex form of the Morlet wavelet, it has better adaptability and time-frequency localization characteristics. Therefore, this embodiment uses the complex Morlet wavelet as the mother wavelet, and its expression is:

[0079] ;

[0080] Among them, f c is the center frequency of the wavelet, f b is the bandwidth of the wavelet, and the wavelet function selected is cmor100-1, cmor is the complex Morlet wavelet, where 100-1 represents f b -f c, that is, the bandwidth parameter and wavelet center frequency are 100 and 1 respectively. The detailed configuration parameters of CWT transform are shown in Table 1.

[0081] Table 1 Important parameters of CWT

[0082]

[0083] In one embodiment, Figure 2 As shown, it also includes the training steps of improving the RegNet network:

[0084] The vibration signals of the rolling bearings in different failure modes during operation of aircraft engines are collected as training data. The failure modes include normal state, inner race failure, and outer race failure.

[0085] Using the training data to train the improved RegNet network, the trained improved RegNet network is obtained;

[0086] The fault diagnosis step is performed using the trained improved RegNet network.

[0087] In this example, the experimental dataset used is a high-speed aircraft engine rolling bearing vibration signal dataset (DIGR). The experimental setup primarily consists of a high-speed spindle, a test bearing, and a triaxial sensor. The high-speed spindle drives the rotational motion, operating at speeds up to 500 Hz. Accelerometers are used at specific locations in the rolling bearing test setup to acquire raw vibration signals through a multi-channel data acquisition system.

[0088] In this example, each of the three different fault modes is categorized into six fault types and one healthy type based on defect depth. Each sample data set consists of 100,000 data points collected using an accelerometer at a sampling frequency of 51.2 kHz and a sampling time of 10 seconds. The specific fault types and labels are shown in Table 2.

[0089] Table 2 Fault types and labels

[0090]

[0091] The CWT time-frequency images of the seven types of faults after CWT time-frequency conversion are as follows: Figure 3 As shown, Figure 3 Figure (a) is the two-dimensional time-frequency image under normal conditions, and Figures (b) to (g) are the two-dimensional time-frequency images corresponding to the IR45, IR25, IR15, R45, R25, and R15 sample data, respectively.

[0092] After completing the wavelet transform, the CWT time-frequency data was divided into training and test sets in a ratio of 7:3. Next, the RegNet-CMSAM model was trained and continuously fine-tuned to achieve optimal results, providing data support for subsequent deep learning model training and evaluation. This time-frequency domain feature extraction method effectively captures fault information in the signal, providing more accurate and comprehensive input features for rolling bearing fault diagnosis.

[0093] In one embodiment, the channel-multi-scale spatial attention mechanism consists of two parts: channel attention and multi-scale spatial attention. Its structural framework is as follows: Figure 4 , the specific steps include:

[0094] Through adaptive global average pooling, the feature map output by the n-1th convolutional layer is compressed in the spatial dimension to obtain the compressed features;

[0095] The compressed features are restored to the original number of channels through convolution, and the channel features are restored and integrated;

[0096] The Sigmoid activation function is used to normalize the weight value of each channel to the range of [0, 1]. The generated weight is multiplied channel by channel with the feature map output by the n-1th convolutional layer, and finally the feature map after channel attention weighting is obtained.

[0097] In the specific implementation of this embodiment, for channel attention, assuming the input feature map size is C, H, W (C is the number of channels, H is the height, and W is the width), first, adaptive global average pooling is used to compress the feature map in the spatial dimension. A global description vector of size C×1×1 is obtained, each vector containing the global key information of each channel in the entire feature map. The calculation formula of global average pooling is:

[0098] ;

[0099] Here, i and j represent the height index and width index in the spatial dimension of the feature map, respectively. represents the eigenvalue of the cth channel at height i and width j. A 1×1 convolutional layer then reduces the number of channels from C (with a compression ratio of r) to C / r, reducing computational complexity while retaining sufficient expressive power to capture key features. Next, a second 1×1 convolutional layer restores the reduced features to their original number of channels, C, to restore and integrate channel features. Finally, a Sigmoid activation function is used to normalize the weights of each channel to the range [0, 1]. The resulting weights are then multiplied channel-by-channel with the original feature map to produce a feature map with channel attention weighting. This process dynamically adjusts the weights of each channel, allowing the model to focus more on important channel features, suppress unimportant ones, and enhance the network's expressive power.

[0100] In one embodiment, the feature Figure 1 The steps for feature extraction include:

[0101] Use multi-scale depth-wise separable convolution to extract spatial attention features;

[0102] Add the spatial attention features extracted at different scales to obtain multi-scale spatial attention weight values;

[0103] The obtained weights and features Figure 1 Multiply channel by channel to obtain the weighted features of the spatial channels Figure 2 .

[0104] In this embodiment, the depthwise separable convolution operation at each scale includes the following steps:

[0105] Pair Features Figure 1 Performing a depth convolution operation, including: performing a convolution operation on each channel using an independent convolution kernel to extract deep features;

[0106] Performing point convolution operations on the depth features, including: using a 1×1 convolution kernel to perform convolution operations on the depth features to fuse information between channels.

[0107] During the specific implementation of this embodiment, for multi-scale spatial attention, the feature map output by the channel attention mechanism is used as input, and 1×1 and 3×3 dual-scale depth-separable convolutions are used to extract spatial attention features. The convolution operation at each scale uses a depth-separable convolution method.

[0108] Depthwise convolution is a special convolution operation consisting of two main parts: depthwise convolution and pointwise convolution. It decomposes traditional convolution into depthwise convolution and pointwise convolution. The first part, depthwise convolution, splits the convolution kernel into individual channels, performing a convolution operation on each channel separately to extract features. Each input channel is processed using an independent convolution kernel, which is not shared with other channels. This avoids computational redundancy between channels in standard convolution and effectively reduces the amount of computation. Next, the second part, pointwise convolution, convolves the output of the depthwise convolution with a 1×1 convolution kernel to achieve information fusion between channels.

[0109] The spatial information of each channel is independently convolved. After integrating the information between channels through point-by-point convolution, the spatial features extracted at different scales are summed to obtain multi-scale spatial attention weights. These weights are multiplied channel-by-channel with the input feature map to generate a feature map weighted for each spatial channel, capturing important spatial feature information. The network parameters are shown in Table 3.

[0110] Table 3 CMSAM structure framework and network parameters

[0111]

[0112] In one embodiment, the improved RegNet network includes an input backbone network (stem), a main network (body), and an output head network (head). The input backbone network is a convolutional layer with a kernel size of 3×3, 32 kernels, and a stride of 2. The convolutional layer includes batch normalization and ReLU activation functions by default. The output head network consists of a global average pooling layer and a fully connected layer. After convolution fusion, global average pooling, and a softmax layer in the fully connected layer, a mapping relationship between fault type and output is established to achieve fault classification.

[0113] In one embodiment, the main network is the core that determines the network computational complexity and performance. In the main network:

[0114] Each stage's Block includes a main branch and a shortcut branch;

[0115] The main branch adopts group convolution operation;

[0116] The shortcut branch adds the input of the main branch to the output of the last layer of the main branch.

[0117] In this embodiment, the main network includes four stages, namely stage1, stage2, stage3 and stage4;

[0118] The Block shortcut branch in stage1 is not processed, and the resolution of input and output remains consistent;

[0119] The Block shortcut branches in stage2, stage3, and stage4 perform downsampling, and the output resolution is reduced exponentially compared to the input resolution.

[0120] like Figure 5 As shown in the figure, Represents the number of output channels, r represents the resolution of the input features, which can be understood as the height and width of the feature matrix, g is the group width of each group in the group convolution, and b is the bottleneck ratio. By adjusting these parameters, the model parameters can be minimized. Figure 1 As shown in the figure, each stage consists of several stacked blocks, and each stage block contains group convolution and ordinary convolution on the shortcut branch.

[0121] The structure of the Block is shown in the figure. The main branch of the Block structure consists of a convolution kernel with a stride of 1 and a size of 1×1, a group convolution with a stride of 2 and a size of 3×3, and finally a convolution with a stride of 1 and a size of 1×1.

[0122] The block-structured shortcut branch has two modes: The first mode, X Block, with s=1, does no processing on the shortcut branch, maintaining the same input and output resolutions. The second mode, X Block, with s=2, performs downsampling using only a 1×1 convolution with a stride of 2. In this case, the output resolution is half that of the input. This modular structure is referred to as an "X-module," and various combinations of X-modules enable flexible network applications in various scenarios.

[0123] In one embodiment, stage2, stage3, and stage4 have the same structure, and the number increases sequentially, such as the number of stage2 is 1, the number of stage3 is 2, and the number of stage4 is 7. Table 4 shows the structural parameters of RegNet-CMSAM, where Conv is a convolutional layer; GAP represents global average pooling; and FC represents fully connected.

[0124] Table 4 RegNet-CMSAM structural parameters

[0125]

[0126] The following is an example of experimental verification of the effectiveness of the method of the present invention:

[0127] To verify the effectiveness of the proposed model, this experiment compares the proposed method with other lightweight and advanced fault diagnosis methods, including lightweight models RegNet-400mf, EfficientNet-B0, ResNet50, and CNN-BiLSTM.

[0128] Taking into account the network structure, computer hardware level, sample characteristics and size, the experimental parameters during training are configured as follows: batch size is 16; learning rate is 0.0003; optimizer is stochastic gradient descent optimizer; loss function uses cross entropy loss function.

[0129] The fault image sample set was divided into training and test sets in equal proportions and applied to each method under the same conditions. Comparison metrics included model parameter count, floating-point computation load, and fault recognition accuracy. Parameter count refers to the total number of trainable parameters in the network model, and floating-point computation load refers to the number of floating-point operations the model performs during inference. The experimental results, shown in Table 5, show that the model parameter count and floating-point computation load of the proposed method are comparable to those of EfficientNet-B0 and RegNet-400mf, and are 19.12M and 48.13M lower than those of RegNet50 and CNN-BiLSTM. Furthermore, the floating-point computation load was 29,555.69M and 1,133.73M lower than those of RegNet50 and CNN-BiLSTM, respectively. Furthermore, the fault state recognition accuracy of each method shows that the proposed method achieved the highest recognition accuracy, reaching 99.25%, an improvement of 0.25% to 1.5% over the highest and lowest accuracies of the other methods, respectively.

[0130] Table 5 Performance index results of different methods

[0131]

[0132] In order to further verify the reliability of the classification results and better present and visualize the diagnostic results of the method, t-distributed stochastic neighbor embedding (t-SNE) and confusion matrix are introduced to visualize the fault identification status results of the proposed method. The original test data and the feature data processed by the method of the present invention are analyzed by t-SNE visualization. The results are as follows: Figure 6 , where (a) is the t-SNE visualization of the original test data, and (b) is the t-SNE visualization of the data after model processing. The corresponding confusion matrix is as follows Figure 7As shown in the figure, the original features are scattered, while the features extracted from each fault mode using the proposed method are clearly separable. Most fault mode features are distributed in non-overlapping spaces, clustered in tighter groups. The confusion matrix also reveals that only a small number of fault types are misclassified. By comparing the computational complexity, parameter count, and diagnostic performance of each model, the results show that the proposed method achieves the best performance in fault identification accuracy, while minimizing model parameters and floating-point computation.

[0133] The following is an example of verifying the noise robustness of the model of the present invention:

[0134] In practical applications, aircraft engine rolling bearings often operate in highly noisy conditions, and the signals collected by sensors contain a certain amount of noise interference. Therefore, the model's fault recognition performance in noisy environments is particularly important. To evaluate the proposed method's ability to extract fault-sensitive information under different background noise environments and verify its noise-resistant robustness, Gaussian white noise with signal-to-noise ratios ranging from -2 dB to 4 dB was added to the original vibration signals of the DIGR dataset to simulate the various noise interferences present in actual aircraft engine rolling bearing operating conditions. Two-dimensional time-frequency images were input into the fault diagnosis model for training. The four most common performance evaluation metrics, Accuracy, Precision, Recall, and F1-score, were used to evaluate the recognition and classification performance of the method for different fault categories. Their expressions are as follows:

[0135] (1) Accuracy: The proportion of correctly predicted samples among all predicted samples by the model. The formula is as follows:

[0136] ;

[0137] (2) Precision: The percentage of correct predictions in the total samples. The formula is as follows:

[0138] ;

[0139] (3) Recall rate: The probability of a sample being predicted as a positive sample among samples that are actually positive. The formula is as follows:

[0140] ;

[0141] (4) F1 score: The relationship between precision and recall, the formula is as follows:

[0142] ;

[0143] Among them, TP means the sample prediction result is a positive sample and the true sample is a positive sample, FP means the sample prediction result is a positive sample and the true sample is a negative sample, TN means the sample prediction result is a negative sample and the true sample is a negative sample, and FN means the sample prediction result is a negative sample and the true sample is a positive sample.

[0144] The calculated evaluation metrics are shown in Table 6. While the model's accuracy decreases as the signal-to-noise ratio decreases from 4 dB to -2 dB, overall performance remains high. Under strong noise levels ranging from 0 dB to -2 dB, the accuracy reaches 92.75% to 89.75%. These results demonstrate that the model has good fault diagnosis capabilities in strong noise environments and demonstrates strong noise robustness and anti-interference capabilities.

[0145] Table 6 Test evaluation indicators of RegNet-CMSAM at different signal-to-noise ratios

[0146]

[0147] The following is an example of how different attention mechanisms affect RegNet performance:

[0148] To validate the effectiveness of the proposed channel-multiscale spatial attention mechanism in improving the RegNet approach, this experiment compares it with state-of-the-art attention mechanisms in terms of key feature focus, model parameter count, and floating-point computational overhead. The attention mechanisms selected for comparison include the squeeze-and-excitation attention mechanism (SE), the convolutional attention module (CBAM), the global attention mechanism (GAM), and the proposed multiscale spatial attention mechanism with depthwise separable convolutional layers replaced by conventional convolutional layers (CCMSAM). In the experiment, these five attention mechanisms were combined with a RegNet network improved with the ELU activation function and tested for fault type recognition under varying levels of noise in the dataset. The experimental results are shown in Table 7. The comparison of model parameter count and floating-point computational overhead shows that the proposed method's parameter count and floating-point computational overhead are on par with those of methods using SE and CBAM. Compared to CCMSAM and GAM, the proposed method reduces parameter count by 6.38M and 4.35M, respectively, and reduces floating-point computational overhead by 2501.31M and 1742.31M, respectively. The results show that compared with conventional convolution methods, the proposed method has significant advantages in terms of parameter quantity and floating-point calculation amount, and can effectively reduce computational complexity and improve the computational efficiency of the model.

[0149] Table 7 Parameters and floating-point calculations of different models

[0150]

[0151] Fault identification comparison results are as follows Figure 8 As shown in the figure, the fault identification accuracy of the proposed method is generally superior to that of other methods, especially under the conditions of no noise, 4 dB noise, and 2 dB noise, where the accuracy reaches 99.25%, 98.75%, and 97.50%, respectively. The fault identification accuracy under the conditions of strong noise of 0 dB and -2 dB is significantly higher than that of other methods. In summary, the proposed method not only reduces the number of model parameters and floating-point computations, maintaining its lightweight characteristics, but also improves fault identification accuracy and demonstrates good noise robustness.

[0152] In summary, rolling bearings in aircraft engines are key components that affect engine performance and reliability, making efficient fault diagnosis crucial for flight safety. Existing fault diagnosis methods have achieved remarkable results in terms of accuracy, but most suffer from large model sizes, resulting in high computational complexity and cost. This limits their deployment in hardware-constrained devices and real-time applications. To address this issue, this paper proposes a lightweight fault diagnosis method based on an improved RegNet and a channel-multiscale spatial attention mechanism.

[0153] Experimental validation on the DIGR dataset demonstrates that the proposed method improves diagnostic accuracy while effectively reducing the number of model parameters and computational complexity. Specifically, the proposed method achieves an accuracy of 99.25%, an improvement of 0.25% to 1.5% compared to other methods. Furthermore, when noise levels ranging from -2dB to 4dB are added to the DIGR dataset, the recognition accuracy reaches 89.75% to 98.75%, demonstrating its excellent noise robustness.

[0154] Compared with existing attention mechanism methods (such as SE, CBAM, and GAM), the proposed channel-multiscale attention mechanism performs better in feature extraction, while also requiring fewer parameters and floating-point computations, resulting in a more efficient model. In summary, the lightweight model-based fault diagnosis method proposed in this paper can maintain high fault identification accuracy under high-speed and high-noise conditions and exhibits good noise robustness, making it suitable for fault diagnosis of rolling bearings in aircraft engines operating under high-speed and high-noise conditions.

[0155] The above is a detailed introduction to the lightweight model aircraft engine rolling bearing fault diagnosis method provided by the present invention. In this embodiment, specific examples are used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0156] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in this embodiment may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown in this embodiment, but is intended to conform to the widest scope consistent with the principles and novel features disclosed in this embodiment.

Claims

1. A lightweight model of an aircraft engine rolling bearing fault diagnosis method, characterized in that: The steps include: Signal acquisition steps: Collecting vibration signals during the operation of an aircraft engine rolling bearing and converting the vibration signals into a two-dimensional time-frequency image; Improved RegNet network troubleshooting steps: The two-dimensional time-frequency image is input into the improved RegNet network; the improved RegNet network sequentially includes an input backbone network, a main network, and an output head network; the main network includes several stages, each stage includes several stacked blocks, each block sequentially includes n convolutional layers and an Elu activation layer, and a channel-multi-scale spatial attention mechanism is applied before the nth convolutional layer for feature extraction, where n>2; the two-dimensional time-frequency image is sequentially input into the several stages through the input backbone network to obtain feature results, and the output head network receives the feature results, establishes a mapping relationship between the fault type and the feature results, and implements fault classification; Channel-multi-scale spatial attention mechanism feature extraction steps: Based on the channel attention mechanism, feature extraction is performed on the feature map output by the n-1th convolutional layer to obtain a feature map 1 weighted by the channel attention; based on the multi-scale spatial attention mechanism, feature extraction is performed on the feature map 1 to obtain a feature map 2 weighted by the spatial channel; the feature map 2 is input to the nth convolutional layer.

2. The method for diagnosing rolling bearing faults of a lightweight model of an aircraft engine according to claim 1, characterized in that: The signal acquisition step includes: converting the vibration signal into a two-dimensional time-frequency image using continuous wavelet transform.

3. The method for diagnosing rolling bearing faults of a lightweight model of an aircraft engine according to claim 2, characterized in that: The complex Morlet wavelet is used as the mother wavelet of continuous wavelet transform.

4. The method for diagnosing rolling bearing faults of a lightweight model of an aircraft engine according to claim 1, characterized in that: It also includes the training steps for improving the RegNet network: Collect vibration signals of aircraft engine rolling bearings in different fault modes during operation as training data; The failure modes include: normal state, inner ring failure and outer ring failure; Using the training data to train the improved RegNet network to obtain a trained improved RegNet network; The trained improved RegNet network is used to perform a fault diagnosis step.

5. The method for diagnosing rolling bearing faults of a lightweight model of an aircraft engine according to claim 1, characterized in that: The step of extracting features from the feature map output by the n-1th convolutional layer based on the channel attention mechanism includes: Through adaptive global average pooling, the feature map output by the n-1th convolutional layer is compressed in the spatial dimension to obtain the compressed features; The compressed features are restored to the original number of channels through convolution, and the channel features are restored and integrated; The Sigmoid activation function is used to normalize the weight value of each channel to the range of [0, 1]. The generated weight is multiplied channel by channel with the feature map output by the n-1th convolutional layer, and finally the feature map after channel attention weighting is obtained.

6. The method for diagnosing rolling bearing faults of a lightweight model of an aircraft engine according to claim 1, characterized in that: The step of extracting features from the feature map 1 based on the multi-scale spatial attention mechanism includes: Use multi-scale depth-wise separable convolution to extract spatial attention features; Add the spatial attention features extracted at different scales to obtain multi-scale spatial attention weight values; The obtained weight is multiplied by the feature map 1 channel by channel to obtain the feature map 2 weighted for the spatial channel.

7. The method for diagnosing rolling bearing faults of a lightweight model of an aircraft engine according to claim 6, characterized in that: The depthwise separable convolution operation at each scale consists of the following steps: Performing a depth convolution operation on the feature map 1, including: performing a convolution operation on each channel using an independent convolution kernel to extract depth features; Performing a point convolution operation on the depth feature includes: using a 1×1 convolution kernel to perform a convolution operation on the depth feature to fuse information between channels.

8. The method for diagnosing rolling bearing faults of a lightweight model of an aircraft engine according to claim 1, characterized in that: In the subject network: Each stage's Block includes a main branch and a shortcut branch; The main branch adopts group convolution operation; The shortcut branch adds the input of the main branch to the output of the last layer of the main branch.

9. The method for diagnosing rolling bearing faults of a lightweight model of an aircraft engine according to claim 8, characterized in that: The main network includes 4 stages, namely stage1, stage2, stage3 and stage4; The Block shortcut branch in stage 1 is not processed, and the resolution of input and output remains consistent; The Block shortcut branches in stage 2, stage 3, and stage 4 perform downsampling processing, and the output resolution is reduced exponentially compared to the input resolution.

10. The method for diagnosing rolling bearing faults of a lightweight model of an aircraft engine according to claim 9, characterized in that: The numbers of stage2, stage3 and stage4 increase sequentially.

Images